# ErezAgent -- User Guide **Version: v4.94.5** > Your AI Company Operating System. One person. A full AI workforce. A command center that executes for you. This guide is the how-to-use-the-product companion to `docs/UI_GUIDE.md`. UI_GUIDE.md is the terse, screen-by-screen reference (what each screen is, when to use it); this guide walks through workflows, setup, and how features connect across screens. Both are organized around the same 13 sidebar groups, in the same order, so what you read here matches what you see in the left navigation. --- ## Table of Contents **Overview** - [What Is ErezAgent?](#what-is-erezagent) - [Installing and Logging In](#installing-and-logging-in) - [Issuing Commands](#issuing-commands) - [The Approval Flow](#the-approval-flow) - [Roles and Access](#roles-and-access) **The 13 Groups** 1. [Getting Started](#1-getting-started) 2. [Executive](#2-executive) 3. [Command](#3-command) 4. [Strategy & Governance](#4-strategy--governance) 5. [Research & Signals](#5-research--signals) 6. [Knowledge & Learning](#6-knowledge--learning) 7. [Decisions & History](#7-decisions--history) 8. [Cost & Finance](#8-cost--finance) 9. [Revenue & CRM](#9-revenue--crm) 10. [Cloud Ops](#10-cloud-ops) 11. [Governance](#11-governance) 12. [Workspace](#12-workspace) 13. [Platform](#13-platform) **Reference** - [Troubleshooting](#troubleshooting) --- # Overview ## What Is ErezAgent? ErezAgent is not a chatbot. It is a structured command center where you direct a full team of specialized AI agents -- each with a defined role, a daily budget, a trust score, and a precise set of capabilities. You operate at the CEO level: you issue goals and commands in plain English, the system classifies your intent, routes to the right agent (or chain of agents), executes, and stops before any sensitive real-world action until you explicitly approve it. What makes this different from a general AI assistant: - **Every AI action is governed.** Agents cannot exceed their defined permissions (Agent Policies, Group 4). - **Every decision is traceable.** Each agent carries a measured trust score computed from real outcomes, not its job title (Decision Intelligence, Group 7). - **Every cost is visible and capped.** Token spend is tracked per call, per agent, per mission (Token & ROI, Group 8). - **Nothing writes to the world without your approval.** Sending an email, publishing content, deploying code, spending money -- all of it stops for your sign-off first. - **Every AI identity is real and disclosed.** ErezAgent itself, your internal agents, and any customer-facing AI representative can carry a persistent name, avatar, and voice -- but identity never grants authority (Persona Management, Group 6). ## Installing and Logging In The fastest path is the **Web Installer Wizard**: 1. Download the installer for your platform from GitHub Releases. 2. Run it -- a browser opens with a step-by-step setup wizard that handles Docker verification, your identity and optional domain, API key entry, secret generation, and Docker launch. 3. On completion, open the app at the URL the wizard shows you. Alternatively, use the shell installer or a manual `docker compose up`. **First login:** you'll see a trial activation screen (self-hosted evaluation). If you have a purchased license, the owner can paste it at any time in Settings > Platform License > Enter License Key; it is checked first and takes effect immediately, with no restart, and the trial screen no longer appears afterwards. The trial lasts 30 days, and a reminder appears in the app from 14 days before it ends. It is also emailed when an email provider is set up (the installer asks; see ADMIN_GUIDE.md section 15), and sent on Telegram when that is connected. **First-time checklist:** | Step | What to do | |---|---| | 1 | Set your `ANTHROPIC_API_KEY` in `.env` | | 2 | Leave `MODEL_PROVIDER_OVERRIDE` empty (disables mock mode) | | 3 | Rebuild: `docker compose up --build` | | 4 | Connect a first provider in Connectors (Group 13) | | 5 | Open the **Setup Wizard** (Group 1) and work through the checklist | Without step 1-3, agent responses come from a mock provider (zero cost, generic answers) -- the visible sign is the AI Provider check on **System Health** (and in the Launch Center), which reads "mock mode" until a real key is set. **Good to know (v4.91.0):** your Executive Office and chat conversations are private to you, spoken replies from OpenAI start playing about a second and a half in, and project deploys wait for approval. **What you see once logged in:** a left sidebar that opens in **Core** view (ten screens: Command Center, Setup Wizard, Executive Office, Build My Business, Mission Control, Approvals, Ventures, Connectors, Docs and Settings) with an **All screens** toggle for the full list of 13 collapsible groups (click a group header to expand or collapse it; the group containing your current page auto-expands; state persists across reloads); a Venture Switcher above the groups that scopes the Command Center and 3D Brain to one venture; a readiness banner that may appear above the content area listing anything still needed before your company is fully operational, with each item linking straight to the page that resolves it; and a persistent command bar for typing goals and commands to the assistant. ## Issuing Commands You can talk to ErezAgent from several places -- the global command bar, Executive Office chat, the Agent Chat tab (Conversations, Group 12), or voice -- and since v4.35.0 they all run through the same underlying assistant. A conversation you start in one place resumes wherever you pick it back up, and a background task's result lands back in that same conversation once it finishes. **What happens when you type a command:** 1. **Intent classification** identifies what you're asking. 2. **Agent routing** sends it to the right agent based on intent and role hierarchy -- or straight to a specific executive if you name one ("CFO, what's our cash position"). 3. **Budget check** blocks execution if the agent is over its daily token budget. 4. **Memory retrieval** pulls the most relevant prior context via semantic search. 5. **Model call** runs, using tools/skills as needed. 6. **Token tracking and audit logging** happen automatically on every call. 7. **Action creation** -- if the agent proposes a real action, it appears in Action Center. 8. **Approval gate** -- if the action is Medium risk or above, it waits for your decision. A command that genuinely contains two separate requests ("post on LinkedIn and also run a security scan") is split and both are handled, up to 3 requests per message. If the system isn't confident what you meant, it asks a clarifying question -- offering its best guess -- rather than silently running with an assumption. **Finding an opportunity when you don't have an idea yet:** you don't need a business idea to start. Say something like *"find me something that can make $5K/month"* and ErezAgent's market intelligence researches real candidate opportunities using live web/neural search, grades the evidence behind each one, and replies with a short, ranked shortlist -- never more than 3-5 candidates, and never one with no real evidence behind it. Reply "go with #2" to move a specific candidate forward; this always requests a real Executive Board review before anything is committed, and you can ask for a cheap validation experiment before betting real budget on it. **Commands that will trigger an approval:** anything that would send a message, publish content, commit or deploy code, spend money, delete something, or change system configuration. ## The Approval Flow **Nothing sensitive executes without your explicit approval.** This is the central governance mechanism in ErezAgent. | Risk | Examples | Default expiry | |---|---|---| | **Low** | Analysis, drafting, research | Auto-approved | | **Medium** | New content creation, internal record changes | 48 hours | | **High** | Send messages, publish content, external API calls | 24 hours | | **Critical** | Financial transactions, deployments, deletions | 12 hours | Every pending approval (whether shown on Approvals, Action Center, Command Center, or inline in chat) shows a risk badge, the requesting agent, a plain-English summary, time remaining, and an expandable full preview with a rollback plan where one exists. Execution-class approvals (a git push, a docker deploy, a shell command) also show an environment tier, a risk gauge, a real git diff, docker context, relevant open security findings, and a rollback command. Approve or reject from the Approvals page (Group 3), inline from Action Center, Command Center's Live Operations panel, Agent Chat, or your connected Telegram bot (instant push with inline Approve/Reject buttons). Rejecting asks for a reason -- pick a quick reason or type your own. On approval the action executes and moves to History; on rejection it's cancelled and the requesting agent is notified; on expiry the Approve button disables but you can still reject. **Multi-step approval chains** (e.g. "CMO reviews, then CEO approves") are genuinely enforced: a later step can't be approved before the one before it, and rejecting one step cancels the rest of the chain. Only the currently-actionable step is shown to you. Two governance gates sit in front of the approval system itself and block execution outright: an unresolved **critical CVE** blocks commit/deploy actions, and a detected **secret leak** blocks publish/push actions, regardless of who approved what. A rejected or expired approval stays rejected/expired -- it can't be resurrected and executed through a different path. Batch-approving several actions at once from a dashboard is owner-only. ## Roles and Access | Role | Can do | |---|---| | **owner** | Full access -- every screen, Settings, user management, approve/reject anything | | **operator** | Create and edit day-to-day work -- tasks, notes, drafts, conversations, customers, sales leads, growth campaigns -- and view all screens. Cannot approve actions (except a step assigned to a business approver role they hold), execute agents, change Settings, manage users, or use the owner-only controls listed below | | **observer** | Read-only across all screens (enforced everywhere since v4.86.0) -- cannot create, edit, execute agents, or approve actions; chat, voice, and previews still work | **Owner-only controls (v4.87.0):** deleting or cancelling missions, projects, contracts, revenue deals, sales companies, finance subscriptions, budgets, resources, strategy plans, autonomous rules, workflows, agent policies, pentest targets, and custom domains; revoking license API keys; resetting a business-policy circuit breaker; deactivating or disabling connectors; deleting voice profiles; and rotating keys or secrets. An owner-only control is shown **disabled** to an operator or observer with an "Owner role required" tooltip, and if an action is refused it is rolled back and the screen tells you why. Operators keep the everyday controls, including customers, leads, campaigns, departments (soft delete), and cancelling a goal or AI plan. **Billing (v4.89.0):** billing changes -- the billing method, gateway credentials, statements, invoices, payments, webhooks and tokens -- need an owner signed in with a real account. Operators can read billing data; observers and venture staff cannot see billing. The static API key cannot make billing changes. **Also owner-only (v4.88.0):** everything that changes your business model in Build My Business (uploading material, approving or applying a change, rolling back, publishing templates), turning AI analysis of your documents on or off, defining venture staff roles, changing an agent's status, budget, model or enabled setting for your organization, restoring an earlier workflow version, and changing whether a workflow requires approval. Operators and observers can read these screens. Manage users from **Users** (Group 13, owner role only): add a user with username/email/role/initial password, deactivate without deleting their history, or reset a password. Two-factor authentication (TOTP) is available per-user from Settings -- Settings > Two-Factor Authentication, scan the QR code with an authenticator app, confirm with the 6-digit code. Once enabled, login needs your password and a live code. Passwords need at least 12 characters and can't be purely numeric or one of the most common weak passwords. As an owner, changing another user's role or resetting their password now asks you to re-confirm your own current password first. Logging out, or changing a password (yours or another user's), immediately ends every other session signed in under that account -- not just the one you acted from. A "Forgot password?" reset link works once; requesting a new one retires any earlier link you hadn't used yet. --- # 1. Getting Started ## Setup Wizard (v4.46.0) **Route:** `/onboarding` A step-by-step path from a new install to your first measured result, read from your account's real data. It shows the same six steps as the **Start here** panel on Command Center and Executive Office: 1. **Set up your company foundation** -- your company's identity and current priority. 2. **State your first goal** -- say what you want to achieve; a short interview turns it into a brief (see Goal Brief in Group 2). 3. **Connect a tool** -- link a connector so agents can take real actions. 4. **Ask the Executive Board** -- send your goal for a Board review. 5. **Approve your first action** -- review a pending approval. 6. **Record your first measured result** -- a real, measurable outcome. A **Next** card at the top names the one step that matters now, with a **Go** button. A progress bar counts the finished steps ("n of 6"), and each step has an **Open** (or **Review**, once done) button that jumps to the screen that completes it. An **Optional** section links to Persona Management (identity), Agent Network (agents) and Build My Business. The wizard is in the **All screens** sidebar view; in Core view, use the Start here panel. Use it right after your first login, or any time you want a guided path instead of hunting for remaining setup steps individually. --- # 2. Executive The executive group is the high-level view of your AI company -- where you review performance, ask the board for advice, and see the overall state of the organization. Open it from the **Executive** group at the top of the sidebar. ## Executive Office **Route:** `/executive` Your daily command post -- start here every morning. Two things live on this page: a **Morning Briefing** generated automatically each day, and a full conversational interface to the board. ### The Morning Briefing The briefing leads with your organization's real state, not markets and news: missions that have stalled or gone overdue, decisions waiting on your approval, and -- from your second briefing onward -- what changed since the last one. These render as clickable cards, not just narrative sentences; click one to jump straight to the mission or approval. Above the narrative, a **Chief of Staff digest strip** shows five tiles: **Needs You Now** (your most urgent items, including any board-tracked assumption that's just been disproven), **Prepared For You** (things already finished without being asked), **Handled For You** (overnight activity), **Waiting For** (things you owe someone, or they owe you), and **Watch** (what's trending toward a problem, including assumptions currently being tested). Click any tile to jump to the corresponding Focus Mode, Decision Queue, or Personal Tasks view. Below the digest: Today at a Glance (key daily metrics), AI Recommendations (top prioritized actions), Overnight Activity, and Intelligence Sources (AI/tech news, crypto movers, GitHub trending, weather, markets) when the relevant API keys are configured. Click **Regenerate** to force a fresh briefing at any time. ### Talking to the board Type a goal or question in the input box to trigger a full board deliberation -- the CEO, CFO, CMO, COO, CTO, and other executives debate the question (see Objective Center, Group 4, for the full deliberation model) and return a structured recommendation. Since v4.44.0, a reply shows the real agent's configured Persona (avatar and name) instead of a generic bot icon, and a small "consulted: X" note whenever a real delegation happened this turn. Text renders right-to-left automatically when the responding Persona's configured language is Hebrew, Arabic, Persian, or Urdu. When a chat creates a mission, that conversation stays linked to it -- open the mission in Mission Control and click **Continue Conversation** to jump straight back with full history. **Attaching a document to a message:** an attached document shows under your message as a small chip with its name and size; the conversation keeps only that chip, not the text. Before the AI sees the document, identifiers in it (Israeli ID numbers, IBANs, card numbers, phone numbers, email addresses, secrets) are masked. What you type yourself is never changed. In **Local models only** mode (Settings > Data & Privacy) attachments are refused with a message, because chat can reach cloud-based skills. The same applies in the command bar and Agent Chat. **Start Customer Solution:** attach a customer's own requirements document (PDF, Word, Excel, PowerPoint, Markdown, or plain text) with the paperclip button, and a **Start Customer Solution** action appears -- this creates a governed solution-delivery mission and extracts structured requirements (see Mission Control, Group 3, for the full requirements-to-blueprint flow). **Set as Company Mission** (v4.63.0): attach your own strategy document the same way, and a **Set as Company Mission** action appears alongside Start Customer Solution -- this extracts a headline, primary metric, target, deadline, key results, and constraints, and creates a *draft* Company Mission in Company Foundation (Group 4). It never activates automatically -- review and activate it there once you've checked it over, same two-step pattern as everything else in Company Foundation. Owner role only. **Set as Venture Foundation** (v4.64.0): when a venture is currently in focus (scoped via "Drill In" from its Ventures card), a third action appears next to the two above -- extracts that same kind of structured identity, but scoped to this one venture instead of your general company, creating a *draft* Venture Foundation for it. Review and activate it from that venture's card in Ventures. Owner role only. **Executive Conversation Mode** gives you a hands-free, continuous back-and-forth: talk, the system transcribes and responds, speaks the reply aloud, and starts listening again automatically. The Neural Core visualization reflects what's actually happening (listening, thinking, delegating, waiting on your approval); if the board needs your approval mid-conversation, the approval card appears right there. Turn on **barge-in** in Settings > Voice Interaction to interrupt the assistant by speaking over it (off by default), and adjust the silence threshold for turn-taking. The original press-and-release mic button still works for single commands. Speech-to-text now auto-detects the language you're speaking instead of assuming English. Turn on **Proactive Notifications** in the same Settings section to have an active voice session speak an urgent alert to you unprompted (off by default). If a command results in "here's the relevant page," a voice-originated turn now takes you there automatically -- a typed command still shows a button to click. ### Decision Queue and Personal Tasks The **Decision Queue** tab shows the same real pending approvals as Command Center and the Approvals page, with working Approve/Reject buttons that call the real backend. The **Personal Tasks** tab is your own persistent to-do list -- separate from agent missions and business actions. Each task has a title, category (personal/business/family/health), priority, optional due date, and notes; all changes persist to the database. A **Waiting For** section tracks commitments you owe or are owed ("I owe Dana the contract by Friday"), and overdue ones surface automatically in your Attention Feed. Since v4.45.0, a second sub-list -- **Blocked across the workforce** -- sits below your own private Waiting For list (never merged into it, since it's tenant-wide rather than personal): it surfaces overdue human-owned work and the AI work genuinely stalled behind it anywhere in your organization. Humans and AI agents can now own the same kind of work item interchangeably, be marked as waiting on each other, and an AI agent that hits a blocker only a person can resolve can hand that work to them directly -- a real task plus one targeted notification, never a broadcast -- and pick it back up automatically once they finish. Reassigning ownership between a human and an AI agent always goes through the same approval flow as any other change; AI never silently reassigns human work. Ask in plain language: "what's waiting on me," "what's blocked by \," or "who's overloaded." **Personal Goals** live here too -- your own goals, separate from the business Goal Planner (Group 12): add one, set a category and target date, and track it through active/achieved/paused/abandoned. If you share an account with other users, your Personal Goals, Personal Tasks, and Waiting-For commitments are private to you, enforced by the database. ### Business Pulse (v4.60.0) Below the Chief of Staff digest strip, **Business Pulse** continuously answers "what changed, why does it matter, does anything need to happen" -- built entirely from your existing Attention Feed signals, not a second monitoring system. Use the venture selector to see the company-wide picture or focus on one venture. Every item is sorted into one of 7 levels, from purely informational up through Critical, and shown across three tiles you can expand: **Needs You** (a decision, an approval, or something serious enough for the board), **Already Handled** (dispatched automatically because it was safe to), and **Watch** (worth knowing, nothing to do yet). An item's own card shows what changed, why it matters, which venture and mission it belongs to, and -- when relevant -- exactly what ErezAgent already did or what it's waiting on you to decide. When the real cause of a stalled mission is known (v4.66.0) -- for example, an email action that failed because a connector is disconnected -- the item states it plainly and shows a **Fix** button that takes you straight there, instead of several separate, unexplained alerts for the same problem. Dismissing an item here dismisses the same underlying Attention Feed signal (and every related signal, if several were combined into one item). Click a venture's Business Pulse to see its **Health** snapshot -- a daily score across revenue, missions, and customers, refreshed once a day. It's an honest subset of the full Company Health score (Group 2): a handful of domains (workforce, security, memory, connectors, assets) don't yet have real per-venture data to score, and the snapshot says so rather than guessing. ### Start here (v4.67.0, one panel since v4.89.0) A slim **Start here** panel at the top of Executive Office and Command Center shows the single next step that matters most on your way to a first real business outcome -- activate your Company Foundation, state your first goal, connect a capability, send your goal to the Executive Board, or review a pending approval, each with a button that takes you straight there. It shows your progress through the same steps as the Setup Wizard. It disappears on its own once a real, measurable outcome is recorded -- it never claims you've reached value from setup steps alone. There's nothing to configure or dismiss; it reflects your real account state. ### Talking to Executive Office (v4.68.0) You can now say things like "run a readiness check", "why is Venture X blocked", or "summarize this mission" directly in Executive Office chat (or the command bar, or by voice) and get a real, immediate answer built from your actual data -- not a guess. When the answer comes with a next step, a clickable button appears with it. "Switch to Venture X", "approve the deployment", "launch a venture", and "build a mobile app" all work the same way they already did on the command bar. While you're viewing a project (Projects) or a customer (Customer Intelligence), "summarize this project" or "summarize this customer" resolves to whichever one you have open -- no need to name it. A project summary covers real status, health, requirement completion, and deployment state; a customer summary covers real health score, churn risk, and open support cases. If nothing's open, Executive Office asks which one you mean. A new **Explain** button on Business Pulse items and the Getting to First Value indicator asks Executive Office the same question with one click, instead of typing it out. Creating a mission or starting a Board deliberation from chat now asks for your approval first, just like any other consequential action -- you'll see it in your Approvals queue before it actually happens. **Acting on a Business Record from chat or voice (v4.85.0):** you can now say things like "move the Acme record to Review", "request a signature on this contract", "request the missing documents from Acme", "schedule an appointment with Acme next Tuesday", or "send Acme a reminder" -- each one runs through the exact same governed, approval-gated path as clicking the equivalent button on the record's own page, whether you say it, type it, or an external MCP client triggers it. Reassigning a task ("hand this off to Dana") is always proposed for your approval, never done immediately. Ask "what's on my plate" or "what needs attention across my ventures" for a spoken summary of My Work or the Operational Dashboard without leaving the conversation. ### Adding a document and stating a goal (v4.89.0) Attach a document in Executive Office chat with the paperclip and one **Add document** button appears. ErezAgent reads the document and proposes where it belongs -- a goal to shape, a customer solution, business material for Build My Business, the company priority, the company constitution or the venture foundation -- and you can choose another destination and must confirm before anything is created (company priority and constitution drafts are owner only). The three earlier buttons (Start Customer Solution, Set as Company priority, Set as Venture Foundation) are still available under **Advanced**. Extraction tells you when a long document was cut short. You can also state a goal in plain words in chat or by voice ("reach 20,000 USD monthly revenue within 6 months"). Executive Office asks up to five short questions, then offers a button that opens the Goal Brief. If it is not sure you meant to start a goal, it asks "Did you mean ...?" first. ## Goal Brief (v4.89.0) **Route:** `/start` (reached from the Start here panel, from Executive Office, or from Add document; it is not a sidebar entry) A goal brief turns a stated goal into something you can check before anything is created. Open the page to type "What do you want to achieve?" and choose **Shape this goal**, or **Resume** an open brief. A brief has three parts: 1. **Review the brief** -- the goal, the success metric (what is measured, target, unit, direction), the timeframe, the scope (the whole company, a new venture, or an existing venture), constraints and the language (English or Hebrew). Fill what is missing or edit anything, then **Save changes**. 2. **Create drafts** -- tick the drafts you want: a Company priority draft (only if you have none; owner only), a new venture (goes through Approvals), a venture foundation draft, and an objective draft. Drafts are never created active. 3. **Ask the Executive Board** -- **Send to Board** is available once the metric, target and timeframe are set; otherwise **Let the Board propose the metric** sends the brief anyway and records the metric as an assumption the Board will challenge. The Board's answer appears in Objective Center. Operators and owners can edit; observers read. **Discard this brief** closes it. Nothing is saved as a company priority, venture or objective until you confirm. ## Build My Business (v4.88.0) **Route:** `/business-twin` (sidebar: Executive > Build My Business) Show ErezAgent how your business works, and it proposes a model of it, with the evidence behind every part. You review and approve the model, then ErezAgent configures the platform to match: record types, stages, rules, document checklists, customer portal pages, workflows, staff roles and appointment types. Afterwards you change how the business runs by saying so in plain language. Nothing is built, and nothing your customers can see changes, until you approve it. **Who can do what.** The **owner** does everything on this page: uploads material, grants AI-analysis consent, approves or rejects items, approves builds, rolls changes back, and manages templates. **Operators and observers** can open the page and read every step but cannot change anything (controls show "Owner role required"). Venture staff accounts cannot reach it. Pick a venture in the Venture Switcher first; each venture has its own model. ### The flow The page walks through six steps; **History** and **Templates** sit beside them. 1. **Sources.** Create a *bundle* (a named set of files) and drop files into it, or use **Or describe your business** and write a few paragraphs instead. Accepted files: PDF, Word (.docx), Excel (.xlsx), PowerPoint (.pptx), CSV, Markdown, text, email (.eml) and images. Scanned PDFs and photos are read as images. Older formats (.doc, .xls, .ppt) and Outlook .msg files are not supported; save them in the newer format or export the message as .eml. A bundle holds up to 25 files and 200 MB by default (50 MB per file); your administrator can change this. Then **Estimate cost**: you see the expected cost before anything is sent to an AI provider, and above a threshold (USD 5.00 by default) you must confirm it. **Analyze** starts the reading; progress shows per file. 2. **Understanding.** What ErezAgent found, in seven areas: Business, Operations, People, Customer experience, AI, Integrations and Management. It also shows questions it could not answer from your material (your answers become assumptions you can review) and a **value hypothesis**: the measurable value this model is meant to create and the assumptions it depends on. If no measurable value is stated, you are told. 3. **Preview.** What your business would look like once built: items per area, each labeled *AI-assisted*, *Rule-based* or *Needs a human decision*; dependency problems that would block a build (for example a workflow whose record type is missing); whether each connected system is really ready; **what it takes to deliver** each requirement (A: ready now, B: configuration or a matching template, C: extension of an existing capability, D: new development, or Blocked when information is missing); **what your staff will see** per role; and **what your customers will see**. Every field a customer can see needs your explicit confirmation. 4. **Decisions.** For each item or whole section: approve, reject, edit, ask a question, or regenerate. **Approve all** skips items with missing information; those need an answer or an edit first. An item that failed validation stays as a gap and cannot be approved. 5. **Build.** **Prepare the build** creates one change for the approved version and works out its impact from your live data: record types, live records per stage, records that would fail new checks, rules, portal pages, workflows, document checklists, role changes, integrations, what customers will see, risks, and whether each step can be undone. If existing records are affected you choose **Apply to new records only**, **Migrate existing records** (done in batches, with a ledger so it can be reversed) or cancel. A change with significant consequences (a role or permission change, customer exposure, signatures, payments or messages, a step that cannot be undone, migrating existing records, or more than 25 changed items) also gets an advisory Executive Board review. **Submit for approval** creates one approval for you in the Approvals queue; only the owner can approve it, and an approval tied to an older version is refused. After approval the build runs step by step, each step verified by reading the result back; if it is interrupted it resumes, and the emergency kill switch stops it. Connected systems are never switched on automatically. 6. **Operate and explain.** Ask about your business and get answers from the approved model and your live configuration, with the evidence shown: overview, how a process works, who handles something, what happens after a stage, why a document is required, which system sends a message, and what is automated. **Ask in Executive Office** sends the question there. **Change how we do this** takes a plain-language change request and prepares it for review in Build. **History** lists every version and change (who requested it, who approved it, when), the **rollback** option, and **Drift**: if someone changes a built part directly in the system, it is marked as changed outside the Business Twin and you are alerted. Nothing is corrected automatically. A rollback shows exactly what will and will not be restored, and goes through the same approval as a build; record types and workflows are restored as a new version rather than rewriting history. **Templates.** Turn parts of an approved model into a reusable template. Names, contact details, documents, credentials, customers and staff are removed automatically, you review what remains, and you acknowledge it before publishing. Installing a template (yours, or one provided with the platform) creates a draft version for the chosen venture that still goes through Preview and approval. Archiving a template does not affect ventures that already installed it. ### Evidence labels Every item says how sure ErezAgent is and why: | Label | Meaning | |---|---| | **Verified from source** | The exact wording appears in one of your documents (checked automatically, not by AI). | | **Inferred** | Reasoned from your material, not stated outright. | | **Assumption** | A guess for you to confirm. | | **Missing information** | Nothing in your material covers this yet. | ### AI-analysis consent Reading your documents with AI is **off until the owner turns it on** in **Settings > Data & Privacy > AI analysis of your documents** (the same card is shown on the Sources step). Two modes: **Redacted (default)** sends document text to your AI provider with identifiers masked first (Israeli ID numbers, IBANs, payment card numbers, phone numbers, email addresses, secrets such as API keys and tokens); **Local models only** uses locally hosted or privacy-safe models and stops instead of falling back to an external provider. Scanned pages and images cannot be masked, so in local-only mode they need a local vision model. The owner can revoke consent at any time. Consent also covers AI reading of customer documents elsewhere: until it is granted, photographed customer documents go to the Review Queue for a person to handle (reason: *AI analysis needs owner consent*), and a scanned PDF with no text layer is listed there as *Unreadable file*. Features you start yourself on your own documents -- customer requirements, contract analysis, knowledge generation, document ingest, company or venture mission extraction, writing style and meeting notes -- do not need consent, and always mask identifiers before the text is sent (in Local models only mode they use a local model or stop with a message). Documents you attach to an Executive Office, command bar or Agent Chat message are masked the same way before the AI sees them; the text you type yourself is never changed, and in Local models only mode attachments are refused with a message. Other countries' ID formats, bank account numbers that are not IBANs, names and free-text addresses are not detected. Uploaded files are kept for a limited time after analysis (90 days by default) and then deleted; the evidence quotes inside the model stay. ### Talking to Executive Office about it Questions work for everyone with access, and you can ask in Hebrew or English (answers are currently written in English): - "Show me how this business works" / "What is automated?" - "How does our onboarding process work?" / "Who handles document review?" / "What happens after the review stage?" / "Why does this customer need this document?" / "Which system sends this message?" Requests to change things are **owner only**, and only ever prepare a change for your review; they never apply it: - "Self-employed customers now require two additional documents." - "Managers must approve discounts above 10%." - "Add a review stage before completion." / "Operations employees should not see financial analysis." - "Send a WhatsApp reminder after three days of inactivity." - "Build my business" / "I want to upload my business documents" opens this page. Answers are short and voice-friendly, with a button that opens the relevant page. ### What gets built, and where to adjust it The build uses the same screens you already know, so you can review or fine-tune the result there: record types and rules (**Business Records**), document checklists (**Document Collection**), customer portal pages (**Business Records > Customer Portal**), staff roles (**Ventures > Staff > Roles**), workflows (**Workflows**) and appointment types (**Appointments**). Edits you make there directly are flagged as drift in History rather than overwritten. ## Briefing History **Route:** `/briefings` An archive of every morning briefing and board deliberation output, searchable by date. Use it to revisit a past recommendation or track how the organization's thinking has evolved. ## Company Health **Route:** `/company-health` A real-time KPI dashboard rolled into a single weighted score across 8 domains: Revenue (20%), Operations (15%), AI Workforce (15%), Security (10%), Customers (15%), Knowledge (10%), Integrations (5%), and Assets (10% -- the share of your real projects and ventures that are currently healthy). Use it for a weekly review to spot underperforming departments before a board meeting. A **Health Metrics** view alongside the domain scores shows how often the board catches its own contradictions, how often past lessons actually get reused in a new decision, and how accurate the board's own predictions have been over time (a calibration trend). --- # 3. Command This is where execution happens: missions are created, agents are dispatched, actions are approved, and runs are monitored. ## Command Center **Route:** `/` The live operations dashboard -- your starting point for "what is happening right now." Active missions, pending approvals, recent agent activity, and system alerts all surface here in a real-time feed. ### The 3D Organizational Brain The centerpiece is a live digital twin of your organization. Each glowing sphere is a real agent, colored by department and clustered in 3D space by department; a sphere pulses faster and larger while its agent has active work. Solid lines are real relationships -- gold for "this agent owns an active mission," magenta for an active agent-to-agent delegation. Hover a sphere for a tooltip (name, role, department, status, trust level); scroll to zoom. Selecting a venture in the Venture Switcher filters the Brain to that venture's activity. An agent with a configured Persona shows its real display name and avatar here instead of the raw agent name -- kept in sync live, so approving a Persona change (Group 6) updates the Brain immediately. Clicking a sphere opens an inspect panel in place (profile, current mission, a "Continue Conversation" link back to the Executive Office session driving it, recent delegation history) without leaving the page; "View Full Profile" jumps to the full Agent Workspace. A **2D toggle** switches to a classic canvas view with the same click-to-inspect behavior and a department legend. A **Live Operations** panel shows who's executing right now and the most recent delegations -- both showing configured Persona names when present. The **Pending Approvals** panel has real Approve / Reject / Discuss buttons on each item; Discuss opens a quick chat with your chief of staff about that specific approval without leaving the page. An **Assets** panel rolls up every project and venture you operate -- Operated, Healthy, Attention, real Cost, and real Value where attributable, next to Active Missions and Ventures. ## Launch Center **Route:** `/launch` The wizard for starting something new: a mission, a strategic initiative, a new venture, or a major project. Three tabs. ### Marketplace tab -- build a project in one click Pick one of 8 pre-defined business application templates (SaaS CRM Dashboard, Customer Portal, MSP Helpdesk, AI Agency Website, Course Platform, Newsletter Platform, Marketing Site, Internal Tool) and the Project Builder Agent generates, scaffolds, and deploys a complete codebase through a 5-step build flow. The result appears in Projects (Group 12). ### Launch Wizard tab -- 3 steps **Step 1 -- choose your goal type** from 17 blueprints (SaaS Product, Consulting Business, AI Startup, E-commerce/DTC, Marketplace/Platform, Content Creator, Education Business, Newsletter/Media, Creative Agency, Open Source Project, Web3/Crypto, Game Studio, Research/Think Tank, Investment Office, Cloud Engineering, Personal OS, or Custom). Each provisions a matching set of executive/specialist agents and 3-4 missions. **Step 2 -- configure**: name your goal, set an optional revenue target, choose a timeline (3/6/12/24 months). A live preview shows exactly what will be created. **Step 3 -- provision**: an animated log shows each resource being created, followed by your agent grid, mission list, success metrics, and a banner suggesting which connectors to set up next. After provisioning, your missions appear in Mission Control, your goal in the Outcome Engine, and your agents are ready to receive commands. ### Orchestrator tab -- idea to live product Turns a single idea into a live, running product through a tracked 9-step sequence: Brand (name, slogan, palette, landing copy, SEO keywords), Scaffold, Deploy (approval-gated -- once approved, the system deploys to your infrastructure and the live URL is stored automatically), DNS (approval-gated domain routing), Payment (approval-gated processor activation), Email, Analytics, Revenue, and Outcome linkage. Start it from the command bar directly ("launch a SaaS for apartment property management") -- the intent is detected, a venture is created, and the orchestrator starts in one message -- or from the **Start Launch Sequence** form on this tab, where you also choose a **Build Engine** (Auto, or a specific model for frontend/full-stack work) that stays consistent for that venture's future build/fix work. **Deployment Provider (v4.58.0):** the launch form also has a real, explicit **Deployment Provider** choice -- Local, Vercel, Railway, Fly.io, or Hetzner (SSH) -- defaulting to Local. Choosing Railway or Fly.io reveals an optional Custom Domain field with a DNS mode of Automated or "I'll configure DNS myself"; in manual mode, you confirm once you've added the DNS records shown to you. Once running, approve each gate as it reaches Approvals and watch step progress update live. If a step fails partway through, the **Saga Rollback Engine** automatically reverses completed steps in reverse order to clean up orphaned infrastructure (for example, deleting an orphaned server if DNS setup then fails), retrying each step up to 3 times before rolling back. If a rollback step itself fails, the resource is recorded as dangling and an Attention Feed alert lets you resolve it manually. ## Agent Network **Route:** `/agents` A centered department tree of your specialized agent workforce. Executives sit at the top; each department branches below in its own color, with team members rolled into avatar stacks. A gold crown badge marks a department's real, admin-declared head. An agent with a configured Persona shows its real avatar (a corner badge on its department icon, or in place of initials in a team stack) instead of just initials, shown only when one is configured. Clicking an agent opens an in-place inspector (Persona header when configured, status, role, trust level, skill count, performance tier, a "Set as Department Head" control) without leaving the page. Search by name/role and filter by department, status, or tier. **Your own workforce settings (v4.88.0):** the core agents are shared starting points, but each organization has its own settings for them -- enabled or not, status, daily token budget, default model, an optional prompt override and department. Changing them affects only your organization, never another one on the same installation, and is owner-only. In the agent's **Settings** dialog, tick **No limit** to remove the daily token budget, and use the checkbox to remove a custom prompt override and go back to the template instructions (both ask you to confirm). Removing an agent from a department takes effect for your organization only. A prompt or permission change still goes through an approved improvement proposal. **Agent Workspace** (click "View Full Profile") shows per-agent detail: role, model assignment, daily token budget, trust score, execution history, skill invocations, and prompt-version history with rollback controls. ## Live Task Runs *(Formerly "Agent Tasks" -- renamed v4.46.0 for clarity.)* **Route:** `/agent-tasks` Give an agent a task in plain words and the agent plans it into steps. - **Safe steps run through the agent**, one after another, and each keeps its real output (for example the drafted email). Later steps are given the results of the earlier ones. - **A step that sends, publishes or changes something needs your approval**; the plan's own "safe" label does not skip this. When a connector is authorised for it (for example a connected email account), the step becomes an approval. An email shows its recipient, subject and text, and goes to an address you wrote in the task. When no connector is authorised, the step is shown as **Blocked** with the reason, and nothing is asked of you. - **After you decide**, the run shows the outcome of each approved step: done, failed or rejected. Executing a task waits for its result, which takes a few minutes when the plan has several steps; the screen waits up to 20 minutes and the run carries on even if you leave the page (open the agent's recent runs to see it). Recent runs of the selected agent are listed beside the form with their status, tokens and estimated cost (the total of the model calls the run made). ## Mission Control **Route:** `/missions` The master list of active and completed missions -- a mission is a multi-step objective with assigned agents, a defined goal, a budget, and measurable success criteria. **Continue Conversation:** every mission card created from an Executive Office chat has a button to reopen that exact conversation with full history. **Lifecycle controls (v4.46.0):** an active mission's card now has real **Pause** and **Cancel** buttons; a paused mission shows **Resume**. A mission stalled at the planning stage (rare, when the AI's own plan validation fails) shows a **Replan** button to retry. These call the same real backend endpoints Executive Office already used for its own mission actions -- now reachable from Mission Control directly. **Since v4.49.0**, pausing a mission (or the platform-wide emergency kill switch) also holds back its Launch Orchestrator steps (deploy/DNS/payment), its customer-support auto-reply emails, and its appointment reminders -- not just new step dispatch and already-approved actions. **Owner-only (v4.87.0):** Cancel and delete on a mission are owner-only; operators see them disabled with an "Owner role required" tooltip. Pause and Resume are unchanged. **Pause genuinely stops it:** no new step starts, and any action already approved for that mission but not yet run stays on hold until you hit Resume -- Resume picks everything back up, including whatever was waiting. **A mission has a real owner from creation.** Every new mission is assigned to a matching executive or specialist agent (a product idea to the CPO, a technical goal to the CTO, and so on, falling back to the CEO for anything ambiguous). That agent posts a kickoff message with concrete first steps into the same conversation, then runs a real plan through planning to execution to completion -- its stage and progress bar advance as steps complete, and anything with real risk stops for your approval exactly as it always does. Agents can hand work to each other for real mid-mission; the receiving agent's work shows up in the same mission. ### From a customer's requirements document to a delivered solution If a customer sends you a requirements document, attach it in Executive Office chat and choose **Start Customer Solution**. This opens a **Requirements** panel (also reachable from the mission's card) with four tabs: - **Requirements** -- every extracted requirement, tagged as explicit, inferred, a recommendation, an assumption, an open question, a conflict, or out of scope. Approve, reject, or amend each one; nothing is treated as agreed-upon until you say so. A requirement that touches something security-sensitive (authentication, admin surfaces, credentials, debug endpoints, encryption, PII, payment data) is automatically flagged here so it gets a closer look during review. - **Coverage & Feasibility** -- gaps, contradictions, duplicates, and the systems involved, with a prioritized list of clarification questions ranked by how much the answer would change the plan. Re-run any time. - **Blueprint** -- once you've approved at least one requirement, **Generate blueprint** turns your approved requirements and the coverage analysis into a structured technical plan: architecture, security model, deployment plan, test strategy, and open decisions. Submit it for board review, and once approved a real Approval request appears on the Approvals page like any other decision. Approving moves the mission into delivery; rejecting sends it back to draft with the board's concerns attached. - **Business Twin** -- when what the customer asked for is mostly about how their business runs (stages, documents, roles, portal pages), **Propose a Business Twin** builds a proposed model from your approved requirements, with each requirement sorted by what it takes to deliver: **A** already supported or configurable, **B** covered by a template, **C** an extension of an existing capability, **D** new development. Review it in Build My Business exactly like a model built from documents (same preview, approval and rollback). If every approved requirement is A or B, delivery is configuration only and no standalone application is generated; the Blueprint tab points you to the Twin instead. If some requirements need new development, the blueprint covers only those. Requirements not yet classified are classified when you propose; until every approved requirement is classified, the Blueprint tab works as before. Once the blueprint is approved, a **Build Engine** selector appears next to Scaffold/retry: leave it on Auto or choose a specific model for frontend/full-stack work -- the same choice persists for future fixes on that project. The delivery team can assemble automatically (a project-scoped specialist agent, or a new connector for a system nothing already talks to) but only ever from capabilities already approved elsewhere in the app; anything genuinely new (like a drafted connector) still needs a person to review and switch on. If the customer already has an approved brand identity, design work uses their real colors and assets automatically. **Android delivery:** ask directly in chat ("build an Android app for Acme Corp that lets staff photograph receipts") and the same requirements-to-blueprint pipeline runs, except the Blueprint step is where the real architecture decision happens -- React Native + Expo is the framework this version builds, and the Blueprint reports honestly when an ask needs something this version can't build yet. Once approved, Scaffold/retry generates a real Expo/React Native project (Expo SDK 57, React Native 0.86, React 19; the app runs on Android 7.0+, and its iOS floor is 16.4 if you build it for iOS yourself), and the Customer Asset tab (Projects, Group 12) gains Android-specific evidence: a real signed build (upload a signing keystore, owner-only), real headless-emulator verification including scriptable interaction flows where a requirement names one, an honest store-readiness checklist, and a governed store-submission handoff (approval creates a real record; actual Play Store submission stays a manual, recorded step). ### After delivery: deploy, accept, maintain Once a project built for a customer solution is running, its **Customer Asset** tab (Projects, Group 12) covers real deployment (an SSH Deploy Target connector, or a Docker export / Kubernetes manifest to run yourself), a live count of how many approved requirements have real passing test evidence, and **Formal Acceptance** -- a checklist built from what you've already approved plus a real security-certification check; Accept only succeeds if every requirement shows real passing evidence and the security check isn't blocked, otherwise it asks for an explicit, recorded owner-only waiver reason. **Since v4.49.0**, a certification with open high-severity findings shows as its own distinct "needs owner review" result -- never silently treated as passing -- so it always requires that explicit waiver rather than an unattended pass. If a scanner itself fails to run, certification is blocked with an honest "scan incomplete" reason instead of reading as a clean result. A change request against this project marks the affected requirement's acceptance evidence stale until it's re-verified, so a change never leaves outdated "passed" evidence looking current. ## Action Center **Route:** `/actions` A live feed of high-risk agent actions awaiting your approval before they execute (send an email, publish content, make a financial transaction, deploy code). Each shows its risk level, requesting agent, and exactly what it will do. Check this multiple times a day -- an unapproved queue blocks agents from finishing work. Actions rated Medium risk or above always need your explicit approval; Low-risk actions can be set to auto-approve in Agent Policies (Group 4). **Real edit persistence (v4.46.0):** the pencil icon on any pending action now saves your edited summary to the server -- an edit sticks across a refresh instead of reverting. ## Autonomy Policy Console *(Formerly "Execution Hub" -- renamed v4.46.0 to describe what it actually controls: how much autonomy agents have, not a generic run monitor. For an agent's own task runs, see Live Task Runs above.)* **Route:** `/execution` The monitoring dashboard for autonomy tiers and execution policy across your workforce. ### Autonomy tiers Every governed action carries a default autonomy tier, which a policy can override: | Tier | Behavior | Use when | |---|---|---| | **T0** | Executes immediately, no human involved | Low-risk, trusted, reversible | | **T1** | Async approval, 24h window, non-blocking | Medium-risk | | **T2** | Sync approval, blocking until you decide | High-risk | | **T3** | Never autonomous, always rejected | Critical | Four tabs: **Live Runs** shows every execution run, most recent first, with gated runs at the top carrying real Approve/Reject buttons; a completed run with a declared rollback action can be rolled back with one click, creating a compensating run of its own. **Action Registry** is a searchable catalog of the built-in action set across categories like email, content, CRM, calendar, sales, legal, finance, research, engineering, outreach, and analytics -- each showing its risk level, default tier, whether a rollback is declared, and estimated cost. **Autonomy Policies** is where you set tier overrides by glob pattern (department-scoped or global) -- a handful of global policies (contract.*, payment.*, code.deploy, code.commit, content.delete) always require approval and cannot be deleted. **Analytics** shows total runs, success rate, spend, and a breakdown by status and tier. Every state transition writes to an immutable execution audit log -- submitted, gated, approved, rejected, started, succeeded, failed, rolled back, cancelled -- with the acting agent or user recorded. It cannot be edited or deleted through any endpoint. ## Capability Engine **Route:** `/capability-engine` Two tabs. **Acquisition Engine** is the marketplace for adding new capabilities to your organization -- find, evaluate, and acquire AI tools, APIs, and MCP servers; each moves through an 8-stage lifecycle from discovery to production. **Capability Registry** browses everything your organization already owns: maturity stage, which departments use it, and its performance metrics. Use Acquisition Engine when you want to extend what agents can do (a web search tool, a GitHub integration, digital signatures, video generation, and dozens more categories). Use Registry to see what's already available before acquiring something new -- the board also checks this automatically before generating a capability-gap recommendation, so you're never asked to buy something you already own. **Readiness (v4.59.0):** a sub-tab under Capability Registry. Type in what you need and get back a real, evidence-based answer -- OPERATIONAL, CONFIGURED but not yet connected, DISABLED, UNAVAILABLE, and a few states in between -- backed by the actual connector, skill, agent, and capability records behind it, never a guess. Enter a mission ID to see that mission's real recorded capability dependencies, with each one's freshness -- if a dependency hasn't been re-verified recently, it's flagged instead of assumed current. Mission planning now uses this same check automatically: a mission step that needs something not actually available is flagged with what's missing before the mission runs, not after. ## Approvals **Route:** `/approvals` The dedicated approval queue -- identical function to Action Center's approval section, presented full-page for when you have many to process. Multi-step approval chains are enforced (see Overview above); only the currently-actionable step of a chain is shown. A recommendation from one of the business-operation skills (weekly business review, meeting follow-up, project/customer risk review, overdue-payment/vendor-renewal review) appears here as a `CONNECTOR ACTION`-type approval, same Approve/Reject flow -- approving dispatches the real connector call immediately. A proposed Persona edit or new Persona creation appears here as a `persona.apply`/`persona.batch_create` approval, same governance as everything else, though you can also act on it inline from Persona Management (Group 6) without leaving that page. **Who asked, why, and what it serves (v4.92.0):** each approval lists the agent that asked (name and role), the reason, and the mission and goal the action belongs to, so you can decide without leaving the page. Expand a card for the preview of exactly what will be sent or changed. A gated skill, such as launching an initiative, lists its inputs one per line. Use this page when you have five or more approvals pending and want to process them efficiently, filtering by risk level. ## Attention Feed **Route:** `/attention` A prioritized feed of everything that needs your attention: expiring licenses, failing agents, unapproved high-risk actions, board alerts, system degradation warnings. Every raw event is scored and routed to a tier -- Immediate (act now), Batch (review at your next session), Digest (read when convenient), Suppressed (logged, not surfaced), or Rejected (failed a duplicate-validity check). Immediate-tier signals arrive within about 2 minutes via live update, with a toast banner counting new signals. Use this page when you see a sidebar notification badge but aren't sure which screen it belongs to. --- # 4. Strategy & Governance Where your company's constitution, strategic plans, and top-level policy live -- the rules the rest of the organization operates under. ## Company Foundation **Route:** `/company-foundation` Your company's constitution and company priority (the current goal it is working toward). This content is injected into every board deliberation, so every AI decision is grounded in your organizational identity. **Company Constitution** is your invariant identity: a one-sentence Company Identity statement (the single most important field -- spend time on it), a long-term vision, core values, and **Non-Negotiables** (things the board must never recommend, however good they look on paper -- the board will actively challenge any recommendation that would violate one). Optional fields (target market, value proposition, business model, moat) improve recommendation quality further. **Company priority** (the Active Mission) is your current, specific, measurable, time-bound goal -- a headline, context, a primary metric and target, 3-5 key results, constraints (budget, team size, technology), and a deadline. "Grow the business" is not a mission; "reach $60K ARR by September 30 with 10 paying customers" is. Mark it **Achieved** when complete (preserved in history, set a new one), or **Pivot** to record why direction changed while simultaneously setting the replacement. **Drafts** created from a document via Executive Office's Set as Company Mission button (v4.63.0) show up here even before any mission has ever been active -- each has an **Edit** button to correct its fields and an **Activate** button to make it live. Nothing extracted from a document ever becomes your company priority without this explicit step. On the Constitution tab, **Start from a document** (owner only) drafts a constitution from pasted text or a loaded .txt or .md file; it is saved as a draft for you to review. A **Board Preview** tab shows exactly what the Executive Board reads at the start of every deliberation -- your identity statement first, then the full constitution, then the active mission. If you run more than one business as fully separate accounts, each tenant has a completely independent constitution and mission; the board for one never sees another's foundation data. If instead you're running several ventures under the same account, each one can have its own **Venture Foundation** (see Ventures, Group 3) alongside this Company Foundation, not instead of it. Set this up during initial configuration and update it whenever strategic direction changes -- it is the single highest-leverage thing to get right, since without it the board reasons without knowing who you are or what you're trying to achieve. ## Objective Center **Route:** `/objectives` Manages every initiative -- not just commercial products -- across 8 types: commercial product, internal business tool, personal goal, research, operational automation, knowledge management, investment, and hybrid. Each type carries pre-defined success metrics and a matching executive focus (an investment objective has the CFO weighing risk-adjusted return, for instance). ### Portfolio, Board Room, and Playbooks The **Portfolio** tab is a dual-lane card grid (commercial products linked to ventures on the left, everything else on the right), each card showing a health dot (green ≥70, amber 40-69, red <40), top metrics, and quick actions to snapshot KPIs, run an AI health review, or open the full record. **Open an objective (v4.92.0):** **Open** on a card opens that objective's own page: its target outcome, success metrics with progress, health and the date of the last AI review, the Board's decisions, its metric history, and a **Board review** link to the deliberation behind it. **AI Review** runs from the page, and **Activate** is offered while the objective is a draft (the capability-gap dialog appears there too). The **Create Wizard** runs a full Executive Board Deliberation before anything is committed: state your desired outcome in plain English, the board deliberates, you review the full deliberation and accept or re-run it, choose (or accept the board's recommended) organizational strategy from 9 types (Product-Led Growth, Enterprise Focus, Geographic Expansion, Acquisition, AI Automation, Cost Optimization, Hybrid Approach, Organic Growth, Platform Play), confirm success metrics, and set a timeline. The objective is not created until you explicitly accept the deliberation. The **Board Room** tab holds the full deliberation history -- every executive's Round 1 and Round 2 positions, the CEO's final synthesis, a governance-confidence score, and a CEO Override badge when the CEO's final call diverged from the weighted score. **Re-Deliberate** runs a fresh session and overwrites the record, useful when the outcome text was unclear or new context should change the analysis. Below the deliberation, an **Organizational Context Graph** panel lists the real context items and warnings behind the outcome; it is hidden when nothing relates yet. The **Playbooks** tab accumulates organizational wisdom across repeated deliberations of the same objective type -- recurring risks, typical conditions, success and failure patterns, confidence ranges, and CEO-override patterns -- and automatically injects those patterns into future deliberations of the same type. Since v4.44.0, each executive's position card in the deliberation view shows their configured Persona avatar next to their role badge -- identity only, never affecting the deliberation itself. ### How the board actually deliberates Every objective runs through a structured 3-round deliberation, not a single black-box call: - **Round 0 -- CEO frames the question**, sharing the core strategic question and domain context with all executives. - **Round 1 -- independent positions.** CMO, CFO, COO, and CTO reason in parallel, with no visibility into each other's answers: a position (Proceed / Pause / Proceed with Conditions), confidence, a summary, risks, assumptions, falsifiable assumptions with a metric/target/checkable-by date, and estimated value. - **Round 2 -- challenge round.** Each executive reviews every Round 1 position and must state who they agree with and why, who they disagree with and why, and whether their own position changed. - **Round 3 -- CEO synthesis.** The CEO resolves conflicts, names who they sided with, and produces a final decision with conditions (if applicable), falsifiable assumptions, and a CEO Override flag if their call contradicts the weighted governance score (CEO 30%, CFO 20%, CTO 20%, COO 15%, CMO 15%). Executives reason from your organization's real live state, not the abstract: the CFO sees real assets/liabilities/budget counts, the CTO sees open security findings, the CMO sees real pipeline and customer counts, the COO sees real mission and approval backlog -- automatically, with no configuration needed. A Board review also shows what the executives could not settle: an executive that lacks the information to judge shows **Abstained**, with the reason and the missing information; open questions are listed on its card; the CEO synthesis notes when every executive agreed with overlapping reasoning ("Treat as one view, not independent confirmation") and lists the **evidence the Board asked for**. Falsifiable assumptions aren't hopes; they're commitments the system checks. Every objective's own assumptions are checked automatically against its real progress on a schedule -- when one is confirmed or disproven, ErezAgent writes a Strategic Lesson (Group 7) and carries it into future related decisions on that objective, even across different wording. Any objective can opt in to this automatic tracking; revenue objectives can opt in out of the box using your real closed-deal totals. ### Scenarios tab Frame a real strategic question -- "Should we allocate the next $100K to Growth or Product?" -- as 2-4 real competing options instead of one free-text proposal: 1. **Create a scenario** with a title and the real question. 2. **Add 2-4 options**, each with a description and optional capital requested; link an option to a real capital-plan bucket or an existing engineering project. 3. **Evaluate each option** deterministically -- no AI call -- using real capital-bucket headroom, the linked objective's health-score baseline, and a real capability-readiness check. Confidence starts `low` and only rises as you add grounded assumptions. 4. **Ground assumptions** by typing one and clicking **Ground via Research** to have the Research Center (Group 5) find real supporting evidence, or record one from your own internal data. 5. **Request Board Deliberation** to have the real Executive Board compare the options directly and rank them. 6. **Request Commit Review** once an option is evaluated and deliberated -- requires at least one grounded assumption, and creates a real Approval. 7. **Approve** on the Approvals page marks the chosen option committed and every other option in the scenario automatically rejected. Use the Scenarios tab specifically when you're choosing between real alternatives, not just tracking a single plan. ## Strategic Plans **Route:** `/strategy` The library of AI-generated strategic plans -- structured plans built from your goals, broken into phases, milestones, and task sequences. Review and approve plans before they execute; each shows its rationale, risk assessment, and expected outcomes. ## Autonomy Score **Route:** `/autonomy` A single composite score (0-100) measuring how independently your AI workforce operates -- what fraction of decisions are made and executed without your manual intervention, across Engineering, Marketing, Finance, Sales, and Operations, with trend over time and a "how to improve" checklist. Use it as a quick health check on whether agents are executing effectively or constantly waiting on you. ## Autonomous Rules **Route:** `/autonomous-rules` The rulebook for autonomous behavior: define conditions and actions ("if the customer health score drops below 40, trigger a check-in email"), operating at any risk tier. Use this when you want agents to react to events automatically without a manual trigger each time. A **Learning** tab measures which rules are actually effective by matching them against completed attribution chains from the last 90 days, updating success/failure counts and a confidence score (0-100) per rule -- green ≥60 performing well, amber 40-59 mixed, red <40 needs adjustment, with AI-generated adjustment notes for low-confidence rules where auto-adjustment is enabled. Runs automatically weekly, or trigger it manually from the Learning tab. ## Operating Doctrine (v4.60.0) **Route:** `/doctrine` Where you define *how* a business, venture, department, agent, or mission should operate -- communication tone, minimum evidence requirements, research requirements, publishing restrictions, spending boundaries -- without any code change. Pick a scope (company, a specific venture, a department, an agent, or a mission), a rule name, and a value; a more specific scope always wins over a broader one when both set the same rule, and every rule shown tells you exactly which scope it's coming from. A small set of security- and approval-critical rule names -- the minimum security floor, whether approval is required, autonomy tier floors, and similar -- can never be set here, at any scope, by anyone. If a new rule's value would conflict with a stricter spending-related limit already set at a broader scope, the page warns you before you save and blocks the save outright for that specific conflict; other kinds of conflicts are shown as a warning but don't block you. Every rule is versioned -- open its history to see every prior value and roll back to any of them with one click. Agents automatically factor in any doctrine that applies to them when they respond, and the Executive Board factors in company-wide doctrine during a deliberation -- you don't need to do anything extra for a configured rule to take effect. ## Agent Policies **Route:** `/policies` The permission framework for agent behavior: what each agent can do autonomously, which connectors it can reach, what risk tier it can self-approve, and what topics are out of scope. Set this up initially and revisit whenever you onboard a new agent or change a trust tier. This is the only place authority is ever granted -- a Persona (Group 6) never touches it. The **Send Emails** / **Use Internet** / **Touch Customers** toggles are enforced the moment an agent tries to act, not just recorded as a setting. **Read CRM**, **Create Deals**, **Send WhatsApp**, **Spawn Sub-Agents** and **Modify Agents** are enforced the same way (since v4.93.0) and start ticked on a new policy. If an agent reaches its daily token budget, the message names the agent and says an owner can raise the budget under Agents > Edit settings. ## Executive Governance **Route:** `/governance` The record of every governance event the board has triggered: health degradations, budget overruns, risk alerts, strategic pivots -- each showing the board's response and whether it was resolved. Use it to review how well the board has been performing its oversight role. --- # 5. Research & Signals Everything that scans the outside world -- and your own product and security posture -- for signal worth acting on. ## Research Center **Route:** `/research` The AI research workspace. Submit a question and the Research Agent produces sourced answers using live web/neural search; every research output is saved and searchable. Use it whenever you need market intelligence, competitor research, or factual background before a decision. ## Opportunity Radar *(Formerly "Value Hunter" -- renamed v4.46.0 to describe what it actually does.)* **Route:** `/value-hunter` An automated opportunity scanner that runs 6 scanner types -- market gaps, pricing optimization, partnership signals, competitor weaknesses, automation ROI, and funding signals -- surfacing actionable opportunities ranked by estimated value. Each is scored with an expected value delta (USD), a confidence level, a cost estimate, time to value, and a computed ROI multiple. Scanners run automatically on a daily schedule; trigger an immediate scan from the Configure tab, and adjust the minimum confidence and minimum expected-value thresholds there if too many low-signal opportunities appear. On any discovered opportunity: **Approve** convenes the Executive Board for deliberation, **Reject** dismisses it, and **Convert to Objective** promotes an approved opportunity into a full strategic objective in Objective Center. Review weekly -- Opportunity Radar surfaces things you wouldn't think to look for yourself. Opportunities that came from asking Executive Office to find you a business idea (**v4.71.0**) show up here too, under a **Market Opportunity** label. Approving one submits it for the same Executive Board review every other opportunity gets -- it stays in **Reviewing** until the Board decides, not instantly approved. If a discovery request produces two ideas that turn out to be the same opportunity worded differently, only the stronger evidence-backed one appears; the other is kept linked to it rather than shown separately. ## Digital Experience **Route:** `/digital-experience` The Chief Experience Officer's dashboard for every digital property you operate -- website, app, landing pages, documentation, social profiles, email campaigns, APIs. Each asset shows a health score (0-100), business-impact metrics where set (revenue influenced, monthly visitors, conversion rate, bounce rate, SEO score, page speed), and a **Generate Recommendations** button that produces 3-5 scored improvement recommendations grounded in the asset's current metrics and your active organizational strategy. When any asset's health falls below 50, the Executive Board convenes automatically with a governance trigger. A digital asset can be linked to a venture, showing a "Linked to venture" badge. Check this after every release, and manage a product or digital service through it continuously. ## Product Assurance **Route:** `/product-assurance` Automated API health monitoring, defect tracking, and formal release certification. **Auto-Discover** scans your OpenAPI spec to build a coverage map; **Browser Scan** drives a headless browser through your UI routes capturing screenshots and console errors; **HTTP Scan** checks every known route's status and latency. Defects carry a severity and type, classified automatically. A **Certification** reviews open defects and returns Certified, Conditionally Certified, or Blocked -- with a forensic record (git commit, app version, coverage percentage, screenshot count, environment) attached. Run this before shipping anything; a release cannot be certified with critical defects open. ## Security Assurance **Route:** `/security-assurance` Three tabs. **Assurance Division** manages security scanning (Bandit, pip-audit, npm-audit, secret scanning, Dockerfile analysis, passive OWASP checks), governed remediation proposals, and independent re-verification of a fix before a finding is closed. **Security Alerts** is the live findings feed, filterable by severity, with Acknowledge/Resolve/False-Positive controls and a security score (100 minus severity-weighted deductions per open finding). **Portfolio** shows a Security Passport per protected asset -- your own products plus ErezAgent itself. A Security Passport is one of six statuses: **PASS** (checks ran recently, nothing open), **WARNING** (something open, nothing critical), **FAILED** (a critical or high-severity problem is open), **NOT_TESTED** (not scanned), **NOT_APPLICABLE** (this check doesn't apply here), or **UNKNOWN**. The distinction that matters most: **NOT_TESTED is never shown as PASS** -- "we haven't looked yet" and "we looked and it's clean" stay visually and semantically distinct, so an unscanned product never gets mistaken for a secure one. Click into a card for the full breakdown -- threat modeling, dependency/secret/container scanning, SBOM and release signing, adversarial testing, open findings, and a trend history. An **emergency Containment** control lets you block a specific route/method combination, optionally tied to an incident ID and an auto-expiry time -- health, auth, and metrics paths are always exempt. Governed containment actions (revoke a credential, disable a user, rotate a secret, roll back a deployment) all require your approval. Security exceptions are a real, governed, time-limited grant with a reason, approver, and expiry -- never a finding simply marked away with no record. Every real project build automatically runs the same security scan against that project's own generated code -- see Formal Acceptance and Publish on the Projects page (Group 12) for where this gates, with an explicit, recorded override available. Never ship with an open critical CVE or an unresolved secret leak; both are blocked at the execution layer, not just flagged here. ## Intelligence Graph **Route:** `/intelligence-graph` A causal pattern graph built from your organization's historical decisions and outcomes -- surfacing hidden relationships between actions and results: what causes revenue to grow, what triggers agent errors, what predicts deal success. Use it for a monthly strategic review, asking "why" questions directly. Board deliberations draw on the same underlying organizational context automatically: existing strategic objectives aligned to a new goal, past decisions on related topics, attribution chains that proved or disproved similar strategies, and relevant strategic lessons. A confidence score accompanies every injected item -- 80%+ is proven via a direct reference, 40-79% is inferred, below 40% appears only as a warning and is never injected into board context. A relationship the system discovers automatically can be reviewed by a person: **Confirm** validates it, **Reject** invalidates it with a reason, **Supersede** links it to a newer, better relationship. Every review action is tied to your authenticated identity. ## Contract Intelligence **Route:** `/contracts` Upload contracts and the AI extracts key terms, flags risks, and suggests questions to ask before signing. **+ Add Contract** creates a new contract record directly on this page. On any contract, **AI Review** lets you choose which review to run -- vendor/reseller agreement, contractor IP-assignment clause, OSS license compliance, or trademark/business-name clearance -- with the result shown inline. The Legal executive behind these reviews also drafts documents on request from anywhere in chat: a mutual NDA, a SaaS Master Services Agreement, SaaS Terms of Service, and more, for any jurisdiction. Use this page before signing any vendor, customer, or employment agreement. --- # 6. Knowledge & Learning Everything your organization has learned, remembers, and can teach its agents to do. ## Memory & Vault **Route:** `/memory` The semantic memory system -- every significant agent output, every uploaded document, every synced Obsidian note is indexed here as searchable vectors. Type a query to search directly; results show similarity percentage, memory type, source, and age, with more recent memories weighted above older ones of similar relevance (a 90-day half-life). **Uploading documents:** PDF, Word, Excel, PowerPoint, CSV, and Markdown/plain text are all accepted (50 MB limit). Each is converted to clean Markdown, saved to your Obsidian vault, indexed as a searchable memory with an embedding, and deduplicated against near-identical prior uploads. **Import Evernote** accepts a `.enex` export and converts every note inside it the same way, in one batch. Open any Obsidian note in the preview panel and click **Summarize** for a real AI summary, or choose a target language and click **Translate** -- both run directly from the preview panel. **Knowledge Graph tab** renders your entire memory space as a real-time neural network visualization -- 12 cluster lobes by domain (Revenue, Engineering, Growth, Security, Finance, Customers, Research, Knowledge, Decisions, Projects, Intelligence, Personal), node shape by memory type, and color by recency ("heat"). Hover for detail, click to select and highlight connections, drag to reposition, scroll to zoom. **Business DNA tab** shows organizational patterns synthesized from your decisions, campaigns, finance, and outcomes -- operational, revenue, marketing, technical, leadership categories, each with a confidence bar and validation count. Click **Run Synthesis** to trigger an immediate pass, or **Validate** a pattern to confirm it's accurate (raises its confidence and validation count). Synthesis also runs automatically on a weekly schedule. Use Memory & Vault whenever you want to find what the system "knows" about a topic, or to audit what agents are drawing on when they make decisions. ## Knowledge Center **Route:** `/knowledge` The Obsidian-based knowledge base viewer -- browse and search vault notes synced into the system. Sync is two-way: approved board decisions and completed missions automatically write readable notes back into your vault, a manual **Sync Now** button triggers an on-demand pull, and a background job keeps it current every few hours regardless. Use this to reference your own notes and research, or to browse a readable trail of past decisions and mission outcomes the system wrote for you. The **Generate tab** produces runbooks, onboarding material, and (with the **Write Content** option) general written content on request. Anything generated here -- including runbooks and onboarding docs -- can be saved straight into your Obsidian vault with one click via **Save to Obsidian**. Beyond browsing, this is also the organizational reasoning engine -- gap detection, contradiction analysis, and duplicate-memory detection across everything the system knows. ## Skills Registry **Route:** `/skills` The catalog of named skills available to agents -- currently 250+ (including 42 gated for an Israel-market deployment, only active when your deployment or a venture is scoped to Israel). Each skill has a defined input/output contract, permission requirements, and a usage log; skills are the building blocks of agent capability. An agent can call a skill mid-response on its own judgment, still approval-gated for anything high-risk. **Real action controls (v4.46.0):** any skill's card now expands into real controls, not just catalog stats -- **Evaluate** runs the Skill Creator eval harness against its current version, **Regression Check** replays recent real invocations against the current definition to catch a quality drop, and **Invoke** lets you run it directly against a chosen agent with your own JSON inputs. A **Propose New Skill** button at the top starts the governed create/evaluate/improve flow -- pick a category, describe the gap -- with the resulting proposal reviewed on Self-Improvement below, same as a skill improvement proposal generated automatically. When the nightly quality review flags a skill, a **"Review in Self-Improve ›"** button on its warning banner (v4.48.0) takes you straight to the pending revision instead of leaving you to find it. Use Skills Registry when you want to understand exactly what an agent can do, or when debugging why an agent couldn't complete a task. ## Persona Management (v4.46.0) **Route:** `/personas` The central, self-service home for every AI identity in your organization: ErezAgent's own persona, every internal agent's, and any customer-facing **External Representative** persona -- name, avatar, tone, voice, language, and communication style. ### What a Persona is -- and is not A Persona is presentation only: `display_name`, `title`, avatar, tone, formality, communication style, technical depth, greeting/closing style, approved/prohibited terminology, primary language, and a voice pulled from your existing Settings > Voice Profiles. **It never grants an agent any new capability, connector access, or risk tier** -- authority lives entirely in Agent Policies (Group 4), and that separation is enforced at the code level, not just by convention: the permission-gating logic that decides what an agent is allowed to do never reads Persona data, and a dedicated test suite verifies this holds by deliberately trying to violate it. A Persona's tone and preferred terminology do shape how an agent phrases a response it was already going to give -- identity only, never what it's allowed to do. ### Using the page Personas are grouped by scope: **System** (ErezAgent itself), **Internal Agents**, and **External Representatives** (customer-facing, scoped to a venture). Each card expands in place into a full edit form covering every editable field, plus a real avatar upload -- drop your own image or have one generated; either way, you review it before it appears anywhere. A **"Configure My Whole Team"** button proposes an identity for every agent that doesn't have one yet, in one batch -- or, since v4.48.0, pick a single agent from the dropdown next to it and click **"Add This One"** to draft just that agent's identity on its own. Either way, nothing is decided for you until you approve the draft. Every change here is a real, governed proposal -- nothing applies directly. Submitting shows a **Pending Approval** state right on the card with inline Approve/Reject buttons (the same approval endpoints Approvals uses), so a self-service edit never has to leave this page to be approved. Once approved, the change is live everywhere a Persona is shown -- Agent Network, Command Center's 3D/2D Brain, chat, the entity inspector, and (for an External Representative) your Support and Booking Portals -- with no page refresh needed. The "Configure My Whole Team" batch never overwrites a Persona you've already customized, and approving the batch only creates drafts -- each individual identity still needs its own separate approval to go live. ### Setting one up in plain language You don't have to use this page directly -- ask anywhere you already talk to ErezAgent: *"give my Security Agent a professional identity,"* or *"create a Hebrew and English support representative for Venture X."* ErezAgent extracts the requested fields and routes to the right proposal type automatically, landing on the Approvals page (or right here, inline) exactly the same way. ### What you'll notice once approved An agent with a configured identity shows its real avatar in Agent Network instead of a department-colored icon, and its real name/avatar on the Command Center Brain (both 3D and 2D). An executive with a configured identity shows an avatar next to its role badge on the Executive Board. In chat, you see who you're really talking to, and -- only when it genuinely happens -- a small "consulted: X" note for a real delegation this turn. Agent Network's live delegation lines show which real identities are talking to each other. If you've configured a Support representative for a venture, your customers see that name and avatar on the Support and Booking Portals instead of the generic "Support" label. Text renders right-to-left automatically in chat, portals, and system notifications whenever the responding Persona's configured language is Hebrew, Arabic, Persian, or Urdu. Use Persona Management whenever you want to give ErezAgent or an agent a real name and voice instead of a generic label, or configure how a customer-facing AI representative presents itself -- and it's the fastest way to see at a glance which agents still have no identity configured. ## Self-Improvement **Route:** `/self-improvement` The self-improvement loop dashboard. On a scheduled review, the system evaluates its own performance across five categories -- cost optimization (agents with high spend where a cheaper model would suffice), prompt optimization (agents with low decision accuracy), budget review (agents hitting their daily ceiling), security review (findings open too long), and architecture review (stale connectors) -- and writes a proposal for each finding: current state, proposed change, expected impact. **Approve** applies the change immediately (a new prompt version is created and activated, for a prompt proposal); **Reject** dismisses it. The system cannot modify its own prompts, budgets, or model assignments without an approved proposal on record -- this rule holds even for background workers, with every applied change logged in the Audit Log (Group 11) and every prompt change roll-backable from the Agent Workspace. This list also includes Skill Creator proposals (a new or improved skill, same Approve/Reject flow) and Customer Solution Delivery's dynamic delivery-org proposals (a project-scoped agent identity, or a new connector module) -- both of the latter are refused automatically before ever reaching this page if they'd request a capability outside what's already approved elsewhere in the app. --- # 7. Decisions & History The institutional memory of the organization -- what happened, why, and what was learned. ## Chronicle **Route:** `/chronicle` Two tabs: **Chronicle** is a filterable event log of every system event (agent runs, approvals, errors, governance triggers); **Timeline** renders the same events as a visual chronological timeline, useful during incident investigation. Open any entry (a mission or venture launch) and click **Replay** to step through its recorded events one at a time with Prev/Next, or let it autoplay -- useful for walking back through exactly how a launch or mission unfolded. ## Decision Intelligence **Route:** `/decisions` The agent trust and decision-accuracy system. Every agent carries a trust score per decision category (Revenue, Technical, Security, Communication, Strategic, Research Quality, Execution Accuracy): **Trusted** (≥80% accuracy) decisions are presented directly to you; **Cautious** (60-79%) are flagged with a caution note; **Unreliable** (<60%) requires a mandatory second-agent review before a decision in that category reaches you at all. Significant decisions create a `decision_record` with a verification window; click **Verify** on any card, select the outcome (Correct / Incorrect / Partially Correct / Not Applicable), and trust scores update automatically. Overdue verifications carry an **OVERDUE** badge and surface in the Attention Feed. An analytics panel shows overall accuracy, a 30-day trend, and the most/least accurate agents. Use this page when an agent seems to be making bad calls -- and periodically anyway, since verifying decisions directly improves future routing accuracy. ## Outcome Engine **Route:** `/outcomes` Links every closed deal, confirmed payment, and completed mission to your primary goal. A progress bar shows current value against target; contributing nodes (deals and missions) are ranked by potential contribution and clickable; active blockers (a stalled deal, a blocked mission, a thin pipeline, an open critical security gate) are auto-detected and severity-sorted. A "What Moved the Needle" timeline shows the chronological attribution trail, with a sparkline of cumulative progress over time. Use this for an end-of-quarter accountability review against your stated goals. ## Decision Outcomes *(Formerly "Attribution.")* **Route:** `/attribution` Causal attribution chains -- which agent actions, decisions, and activities contributed to which outcomes (revenue, cost savings, churn events), with a full audit trail of cause and effect. Use this to see which agent activities are actually moving the needle, or to justify a past decision to yourself retrospectively. ## Strategic Lessons **Route:** `/strategic-lessons` The institutional memory for lessons learned -- every significant decision outcome, good or bad, generates a lesson stored here and referenced in future board deliberations automatically. A summary bar shows total lessons, high-impact count, "what worked" vs "cautionary" counts, and total value captured (the summed USD outcome delta across measured lessons). Each lesson carries an impact badge (High/Medium/Low), a type (what_worked, what_failed, assumption_validated, assumption_invalidated, pattern_discovered, risk_identified), and a domain (revenue, growth, security, customers, workforce, finance, operations, knowledge, general). Expand a card for the full text, the action that produced it, the observed outcome, tags, and a **"Mark Reviewed"** button (v4.48.0) -- click it once you've acted on the lesson or decided it needs nothing further, and it drops off this list. **Extract New Lessons** runs an on-demand extraction pass over recent attribution chains and executive decisions; extraction also runs automatically on a weekly schedule. A lesson generated from one of your objectives' own assumptions shows which objective it traces back to, and how many times it's since been surfaced into a later decision. A question or complaint that recurs across social comments and mentions becomes a written lesson here too, the same way a recurring objective outcome does. Read high-impact lessons weekly, and check cautionary lessons filtered to a matching domain before starting a similar new objective. ## Department Evolution **Route:** `/department-evolution` Two tabs. **Performance** tracks department-level KPI hits and misses -- consecutive misses trigger automatic capability-gap analysis. **Workforce Analytics** shows AI workforce composition, agent utilization rates, and cost-per-department breakdown. Use this for a monthly department review to identify which teams need new capabilities or agent reinforcements. When you expand a capability-gap recommendation, a tool your organization already owns appears with a green "Acquired" badge and a $0 cost label (the board checks the Capability Registry before recommending a purchase); anything genuinely new to acquire shows its estimated monthly cost. --- # 8. Cost & Finance Everything about what your AI organization costs and how well that spend is performing -- collected in one place. ## Token & ROI **Route:** `/tokens` Token usage analytics tied to business outcomes -- which AI activities generate the highest ROI per token spent, broken down by agent, model, and mission, with a 14-day usage chart. Use this when optimizing AI spend, to find the activities where cost-per-outcome is too high. ## Finance Department **Route:** `/finance` The Finance Division dashboard: P&L, budget vs. actuals, cash flow projections, invoices with a real sent → overdue → paid lifecycle (a daily check flips overdue invoices automatically), and CFO recommendations, linked to venture wallets and department budgets. Invoice numbers are issued in order, and an owner records a payment against an invoice (amount, reference, date received, method and an optional evidence note); each payment shows where it came from (processor-confirmed, owner-attested or imported) and a wrong one can be reversed with a reason. These are invoice drafts, not tax invoices. A **Transactions** tab is the debit/credit ledger (manual, Stripe, model-call cost, or automatic) with net position and a manual-entry form. A **Procurement** tab covers vendor subscriptions -- cost, billing cycle, renewal date, category -- with summary cards for total monthly spend and renewals due, and a **Run Reconciliation** action that compares subscription costs, model-call costs, and manual transactions for the current month against expected spend, flagging Balanced / Overspend / Underspend with a historical report archive. Use this for your monthly close: review CFO analysis, approve budget adjustments, check Procurement for renewals due. ## FinOps **Route:** `/finops` Cross-provider cloud cost management -- spend breakdown by provider, waste detection (idle VMs, over-provisioned instances, abandoned volumes with estimated monthly savings), a 3-month cost forecast, and an AI-generated CFO report. Budget alerts fire when a provider exceeds its configured alert threshold, and a cost-spike signal fires when a provider's current-month spend exceeds last month's by more than 20%. Review monthly. ## AI Model Routing & Cost *(Formerly "Intelligence Economy" -- renamed v4.46.0.)* **Route:** `/intelligence-economy` The intelligence spend model -- how AI capabilities are being consumed across departments, procurement decisions, budget allocation for intelligence, and ROI per capability. Use this for a monthly FinOps review of AI spend specifically: which capabilities to expand, which to retire based on ROI. ## AI Spend Intelligence *(Formerly "Value Intelligence" -- renamed v4.46.0.)* **Route:** `/value-intelligence` A CFO-grade savings dashboard, computed from data AI providers already return on every call -- zero additional token overhead. An all-time savings counter only ever goes up; a period selector (7D/30D/90D/12M/All Time) rescopes everything; an Efficiency Score (0-100) combines cache hit rate, routing efficiency, call success rate, compression ratio, and dedup rate. A stacked area chart breaks savings into four types -- cache, routing, compression, dedup -- and an agent ROI ranking table and model spend distribution round out the picture. Savings Opportunities surfaces 2-3 algorithmic recommendations (low cache hit rate, an unnecessarily expensive model, an uncompressed large prompt) with estimated monthly savings. **What-If Simulator:** pick a "from" and "to" model and run a simulation against your actual historical usage data -- returns projected monthly savings/cost change, annual impact, a quality/latency verdict, and a confidence level based on call volume, with zero additional AI calls (pure arithmetic against real records). **AI Spend Forecasting:** projects your spend for the next 30/90/365 days from a linear regression over your actual daily cost history, with a trend direction and a confidence level based on how many days of real data you have. **Department Spend / Chargeback:** if you organize agents into departments (Custom Departments, Group 11), this page shows per-department AI cost, savings earned, and share of overall spend, with a chargeback report (gross cost, savings, net cost) for finance reporting. Use this page whenever you need to justify ErezAgent's own operating cost. --- # 9. Revenue & CRM Everything related to customers, revenue, and commercial operations. ## Sales Department **Route:** `/sales` A Kanban pipeline (Prospect → Contacted → Qualified → Proposal → Closed Won) for managing leads from first contact to closed deal. Each lead card shows a 4-cell BANT grid -- Fit, Urgency, Budget, Authority, each 0-100 -- scored by the CRO agent on request, with the reasoning available on hover. **Outreach** generates a personalized first-touch email and routes it to Approvals; it's only sent once you approve it. Every lead has a collapsible interaction timeline (note, call, meeting, email sent/received, proposal, objection, follow-up) with a positive/neutral/negative outcome marker. A **Companies** strip groups leads by company automatically, and each company badge opens a **Company DNA** panel -- stats, an outcome breakdown, a pipeline breakdown, and an AI-written strategic narrative describing the relationship's depth and potential, refreshable on demand. **AI Prospect** queries live web research for real companies matching a description you type, adding them as leads automatically. A lead can belong to a venture. **Add Lead** has a Venture picker; the **Venture** filter above the board shows one venture's leads, and **AI Prospect** assigns the leads it finds to the venture chosen in that filter. Each lead card has a venture picker, so a lead can be moved to another venture or taken out of one. Leads from the website forms go to the venture whose slug is `erezagent` when your workspace has one. Promoting a lead to a deal keeps its venture. ## Customer Success **Route:** `/customer-success` Customer health monitoring, renewal tracking, and QBR scheduling. Portfolio KPIs at the top (customer count, total ARR, average health, at-risk count, critical count, onboarding-in-progress count); each customer card shows an animated health ring, a churn-risk badge, ARR, NPS, and onboarding status, with an expandable success-task list (categorized onboarding/adoption/renewal/expansion/health, each with status and due date). Filter the board by risk tier. Review weekly -- never let a customer's health score drop to red without a check-in. ## Appointments **Route:** `/appointments` **Venture-aware since v4.42.0, with Appointment Intelligence since v4.43.0.** The admin home for the Appointment & Engagement Engine. Scheduling is a real part of how ErezAgent runs your business, not a separate calendar bolted on -- a booked meeting is tracked the same way a mission or a deal is: what it's for, what was prepared for it, what happened, and what it led to. ### Setting up appointment types An **appointment type** is a reusable template for a kind of meeting -- Sales Discovery, Technical Consultation, Executive Review, Customer Success Review, and others you define. Each type sets who it's for (external, so a customer can book it, or internal-only), duration, notice period and booking window, meeting provider (Google Meet is fully wired; Zoom and Microsoft Teams connect as soon as you provide credentials), and an optional intake questionnaire plus a list of documents you want attached before the meeting. A venture picker in the top bar filters every tab and scopes new types, resources, and availability windows to one venture instead of always being tenant-wide; a venture-scoped, externally-bookable type shows its real public Booking Page link right on its card. You also manage bookable resources (rooms, equipment, demo environments) here. ### The customer booking page Every venture gets a real public booking page at `/book/{venture_id}` -- share it anywhere: your site, an email signature, a proposal. No login required. A customer sees only externally-bookable types, checks real availability (your actual calendar and existing bookings, never a guess), and requests a booking with their email; a one-click verification link (shared with the Customer Support Portal) lets them return later and see their own upcoming and past appointments, with the ability to cancel. A venture can claim its own custom domain for its booking or support page (Ventures, below) so this never has to live on ErezAgent's own domain. ### Booking in plain language Ask directly in chat or the command bar -- *"Book a technical consultation with Acme Corp for next Tuesday afternoon"* -- and ErezAgent infers the venture and appointment type from context, books it when confident, and asks a real clarifying question when a customer name or type is ambiguous rather than guessing. ### Before and after the meeting Ahead of an appointment, ErezAgent prepares a real **Engagement Brief** grounded in your actual history with that customer or project -- prior conversations, open commitments, past decisions -- honest about gaps rather than inventing history to sound complete. Since v4.43.0, the brief also surfaces **Decisions Needed** (what this meeting must actually decide, given your real open commitments) and a **Recommended Meeting Outcome**, both reasoned only from real evidence. You can also set a real **meeting objective** when booking, and ErezAgent tells you afterward -- yes, partially, no, or unclear -- whether it was actually achieved. After the meeting, give ErezAgent your notes (or let it use a real transcript, when your Google account supports it) and it produces a summary, the real decisions made, and action items that become tracked commitments alongside your other open loops. Depending on the meeting type, a confirmed outcome updates the right part of ErezAgent automatically: a decision tied to a project becomes a real timeline entry; a risk raised in *any* meeting -- not only a dedicated customer-success review -- now moves that customer's health and churn tracking; a new opportunity carries the actual text raised in the meeting, not a generic summary; anything needing real follow-up work is proposed as a governed action, still requiring your approval. An unanswered question resurfaces in your next brief for that same customer instead of being forgotten. A follow-up email is drafted for you but **never sent automatically** -- it goes through the normal approval step like every other outbound action. Commitments coming due, and any already overdue, surface as real attention items; a customer relationship that's gone quiet with no pending activity is flagged the same way. Once an outcome is confirmed, you can ask ErezAgent to propose the next meeting -- it drafts type/objective/attendee for your review, but never books it itself. A summary on this page shows which meeting types actually convert to real pipeline value, broken down by venture. ## Revenue Center **Route:** `/revenue` Deal pipeline and revenue events across four tabs: **Deals** (list view, with stage badge, USD/ILS value, probability, weighted value, and Win/Loss reason capture that feeds a recurring-pattern check into Strategic Lessons), **Pipeline** (the same deals as a Kanban board by stage, filterable by venture since v4.87.0), **Attribution** (campaign → deals → revenue rollup, with a venture filter when you manage more than one), and **Billing** (below). The header shows only figures that exist, with no sample data: pipeline and weighted value, won deals, cash collected this month, and contracted recurring revenue (MRR) from active subscriptions only; when your Company priority has a revenue target, a goal bar shows progress toward it. A **Needs Action / At Risk** panel surfaces stalled deals, at-risk customers, and renewals due within 30 days automatically. **Billing (v4.89.0)** is for owners and operators (only owners change things) and has five sections. **Overview** shows, for a month, open drafts, unpaid, overdue and paid invoices, the outstanding balance and cash collected. **Usage** shows what the usage ledger recorded (AI usage and cost, actions, agent runs, capacity), lets an owner add a dated adjustment with a reason, and exports a usage report for your vendor. **Statements & invoices** generates a usage statement from the ledger (prices come only from the unit prices you set; unpriced lines are flagged), drafts and finalizes invoices, and downloads PDF or DOCX documents or sends them by governed email; voiding and credit notes need an owner approval. **Payments** lists and records payments, reverses a wrong one, and imports a bank CSV with a preview. **Integrations** connects your accounting system: signed webhooks, API tokens and exports. Every document is labeled as a usage statement or invoice draft, not a tax invoice. How you are billed or bill others -- metering (you record payments) or a gateway -- is set in Platform Management -> Billing Setup. Beyond manual entry, ErezAgent can run the front end of your sales pipeline for you: point a web form, automation, or chat widget at the acquire endpoint and every submission becomes a deal automatically; the CRO agent qualifies it and drafts a priced quote with a real payment link (Stripe, PayPal, Rapyd, or Pelecard, whichever you've configured and entered in Billing Setup) -- routed to Approvals like any other outreach. Once a customer's contract approaches its end date, a **RENEWAL DUE** badge appears on their Customer 360 record (below) and **Send Renewal** drafts a renewal email automatically; healthy, low-churn-risk customers show an **EXPANSION CANDIDATE** panel for a consultative upsell pitch. Renewal drafting also runs on a daily schedule for anyone entering the window, so nothing depends on you remembering to check -- it still only ever produces a draft. If you already run HubSpot, Salesforce, or Pipedrive, every newly captured lead syncs out to your existing CRM automatically (best-effort; capture in ErezAgent never depends on the sync succeeding). ## Ventures **Route:** `/ventures` Your venture portfolio -- each product or business you're building, with its own wallet, spend gate, and P&L. A wallet enforces daily/monthly spend limits (fail-closed: if the gate check itself errors, the action is blocked, not allowed); live P&L tracks revenue, expenses, gross/net profit, MRR and ARR (from contracted recurring subscriptions only), ROI, burn rate, and runway; and rule-based recommendations surface Scale / Pause / Shutdown / Continue with a rationale and confidence score. **Act on a recommendation directly (v4.46.0):** the Recommendations tab's Scale/Pause/Shutdown cards now have a real **Apply** button next to each recommendation, calling the same pause/resume/shutdown/scale endpoints directly from this page -- Shutdown asks for confirmation first. Each venture card's P&L block sits alongside a separate real "AI cost to build/operate" figure, kept deliberately apart from P&L so your revenue/ROI numbers never silently change. A **Support** button opens that venture's Support Workspace -- the public portal link, branding (logo/color/name), a custom support domain, and an embeddable widget snippet for external sites; a **Self-Service Domains** section lets the venture claim a real custom domain for its Booking or Support page, going live automatically once you point one DNS record at your server, no manual setup on your end. If you've configured an External Representative Persona for the venture (Group 6), the public Support and Booking Portals show that Persona's real name and avatar instead of the generic "Support" label. **Venture Foundation (v4.64.0):** a **Foundation** button opens that venture's own identity -- its own mission, target market, and business model, separate from your general Company Foundation (Group 4). Board deliberations, individual agent actions, and customer-facing communications for this venture now reflect its own Foundation alongside your Company Foundation, not instead of it. Often created automatically: when the Executive Board researches and approves a path that results in this venture being created, a draft Foundation is derived from that same approved recommendation, waiting for you to review and activate -- it never goes live on its own. You can also set or edit it manually, or attach a document in Executive Office chat and click "Set as Venture Foundation" while this venture is in focus. ### Venture lifecycle Each venture moves through a 10-stage lifecycle, shown as a colored pill on its card: idea → building → internal testing → beta → public → growing → revenue positive → scaling → maintenance → sunset (terminal). Not every transition is valid -- the system enforces a defined transition map, and an attempt to skip stages is rejected. Advance manually from the card's dropdown (only valid next stages are offered), or let the background lifecycle sensor auto-advance building → internal testing when a launch's deploy step completes, and flag sunset risk (ROI below -20%, 90+ days active, under 10% budget remaining) as an Immediate attention signal -- advisory only, never an automatic state change. Each card links to a full transition history. ## Venture Commerce **Route:** `/venture-commerce` A venture's own product, subscription, and refund lifecycle for **its own end customers** -- structurally separate from your own Plan & Billing (that's what you pay for ErezAgent; this is what a venture's customers pay the venture). Select a venture from the venture switcher first -- this page is always scoped to one specific venture, never "all ventures." - **Products** -- define what this venture sells. Each product follows the venture's billing method or is **Metering only**. With **Metering**, no checkout is created: the subscription is billed from usage statements and invoice drafts and you record payments when they arrive. With **Stripe** or **PayPal**, the customer completes a checkout through the venture's own connected payment credential. Rapyd and Pelecard are not available here. - **Prices** -- add one or more prices to a product (amount, currency, billing interval, optional trial days). A price's real id is shown right after you create it -- you'll need it to start a subscription checkout. - **Subscriptions** -- start a real checkout for an existing customer: pick the product, the price, the customer, and a payment provider, and a real checkout session opens. Status (trialing, active, past due, canceled) reflects what the provider actually reports. Cancel immediately or at the end of the current period. - **Refunds** -- request a refund the same governed way every other refund in this system works: it creates a real Action and Approval, and only actually processes once you (or another owner) approve it from Approvals -- never an immediate charge reversal. - **Usage** -- define the usage units your venture charges for (label, unit, optional price per unit) and see per-customer usage for a month; your own system reports events against a unit. - **Billing** -- this venture's billing method and gateway credentials (the same screen as Billing Setup, with the venture's own credentials). A method change needs owner approval. Tax on a subscription's checkout is calculated automatically when the customer's country is known, using the same real tax engine (`v4.77.0`) Domain A billing uses -- never a flat guessed rate. ## Launch Infra **Route:** `/launch/infrastructure` Infrastructure provisioning for new ventures -- hosting, domains, databases, and CI/CD for a new product launch. ErezAgent uses a shared infrastructure model: one modest server runs a reverse proxy, and each venture gets its own subdomain, keeping cost minimal as you add ventures. Five tabs cover Providers (Hetzner for hosting, Cloudflare/Namecheap for DNS, Resend for transactional email, PostHog for analytics -- each configured independently via `.env`), Domains, Deployments, Emails, and Analytics. Server provisioning always requires approval regardless of budget; DNS creation and read-only analytics queries don't touch the spend gate. **Railway + Fly.io (v4.58.0):** the Providers tab also has Railway and Fly.io rows with their own **Connect** button -- add your own API token (and, for Fly, an org slug) directly from the UI instead of `.env`. Once connected, they're selectable as a venture's deployment provider from Launch Center. The Deployments table now shows a Region and Est. Cost/mo column for every deployment. ## Customer Intelligence **Route:** `/customers` Your CRM layer -- customer records, interaction history, enrichment data, and AI-generated insight per account. The list shows a color-coded health bar, churn-risk badge, ARR, days since last contact, and open incident count, filterable by tier/health/risk. Selecting a customer opens **Customer 360**: KPI tiles, a contacts tab, an incidents tab, and an **AI Recommendation** section that runs health recomputation, churn-risk classification, and a full AI briefing in sequence on request. The **Contract panel** shows a **RENEWAL DUE** badge and **Send Renewal** button for anyone within 90 days of their contract end date -- clicking it has the CRO agent draft a renewal email with a live payment link and route it through Approvals. A **BRAND & CREATIVE** panel near Expansion has a **Request Brand & Creative Work** button that opens chat pre-filled to generate that customer's own brand identity and an initial creative set -- the same identity a delivery project for that customer automatically pulls into its design context. A **delete** control is available from the Customer 360 header, with a confirmation step. The system also monitors every customer automatically, sending Attention Feed signals when health drops below 40, churn risk goes high/critical, a strategic/standard customer goes 30+ days without contact, or a P1 SLA breach occurs. Review this before any customer conversation, and check contracts approaching renewal proactively rather than waiting for them to lapse. ## CRM Workspace (v4.78.0) **Route:** `/crm` One unified view of a customer, pulling together everything already tracked elsewhere in ErezAgent -- contacts, deals, projects, support cases, subscriptions, activity timeline, and next actions -- instead of hunting across separate pages. Search across customers and contacts, scoped by venture the same way Customer Intelligence is. Selecting a customer opens their workspace: related records grouped by type, next actions with due dates (an overdue one surfaces in the Attention Feed automatically), and a unified activity timeline spanning deals, support cases, documents, commerce, and appointments -- not just the notes you add by hand. A **duplicates** view surfaces AI-proposed matches when two customer or contact records likely refer to the same person or company -- review and merge them yourself; nothing merges automatically. **Import from CSV** lets you bring in a customer list with a dedup preview before anything is written, so you decide row by row whether to create a new record, merge into an existing one, or skip it. See **Customer Operating Intelligence** below for what needs attention with a customer, the AI Customer Brief, and the portfolio-wide attention view. ## Document Collection (v4.78.0) **Route:** from a customer's CRM Workspace or Customer Intelligence page Create a reusable checklist of documents to request from a customer (e.g. "Government ID", "Proof of address") -- either from a saved template or built ad hoc. Each requirement moves through a real status as it progresses: missing, requested, received, needs review, accepted, rejected, or waived, with a full history of who changed what and when. A file arriving does not automatically count as accepted -- that is always a separate, deliberate decision. **Conditional documents (v4.88.0):** in a checklist template, tick **Only when a condition holds** on an item to require it only for some records, for example a second ID only when a "customer type" field is "self-employed", or only once the record reaches a given stage. The checklist updates automatically when the record changes; rules can also add a requirement to a record's checklist. **Requesting documents:** choose a channel (email or WhatsApp -- WhatsApp needs to be connected first, see Connectors) and send a request or reminder. Every send requires your approval first, exactly like any other governed outbound action. A reminder respects each contact's quiet hours and opt-out preference, and the same reminder will not go out twice in a short window even if you ask for it again. **Receiving documents:** a customer can reply by email or WhatsApp with the file attached, or upload it themselves through their own customer portal (the same portal used for support and booking). AI proposes which checklist requirement the file satisfies, showing its confidence -- a confident match advances the requirement automatically, anything uncertain lands in the **Review Queue** for you to resolve by hand. A scanned (non-text) PDF is not auto-classified; it lands in the review queue for a human decision rather than being guessed at. A photographed document (an image) is read by AI only after the owner turns on AI-analysis consent (Settings > Data & Privacy); until then it also lands in the review queue. **Viewing a submitted file (v4.87.0):** each submission in the **Review Queue** has a **View file** button that opens the original the customer sent. Access is limited to your own tenant (and, for venture staff, their own venture), and every view is recorded in the Audit Log. ## Customer Operating Intelligence (v4.79.0) **Route:** from a customer's CRM Workspace, or from Executive Office chat/voice **Needs attention:** each customer's workspace now opens with a **Needs Attention** section -- open findings like churn risk, a renewal approaching, a stalled deal, an overdue document, an SLA breach, a payment issue, or an overdue follow-up. Each finding shows a confidence label (**verified** when it traces directly to a real record, **inferred** when it's an AI judgment) and, where a direct action exists, an **Execute** button that either sends a governed reminder or hands the task to you as a real to-do -- nothing executes without this explicit step. A portfolio-wide **Needs Attention** tab lists every customer across your business needing attention, filterable by venture, finding type, and severity. **AI Customer Brief:** click **AI Brief** from a customer's workspace to generate a governed, versioned summary -- a risk assessment and a recommended focus, grounded only in that customer's real account data. Every section is labeled verified, inferred, or missing, so you always know what's a fact and what's the AI's own read. Ask for it by voice too ("brief me on Acme Corp") -- Executive Office speaks back a one-sentence summary while the full brief is available on screen. **Ask Executive Office about a customer:** you can now ask about a specific customer's findings, next actions, or commercial status by name -- typed or spoken -- even without that customer open on screen ("what does Acme Corp need from us"). You can also ask what's waiting on you versus waiting on the customer, which customers need attention right now, or which renewals are coming up in the next 30 days. **Unified timeline:** the customer workspace's activity timeline now spans every subsystem -- deal stage changes, support case updates, document requests, commerce events, and appointments -- alongside your own manual notes, all in one real, chronological view. ## Licensing Department **Route:** `/licensing-dept` License management for the ErezAgent platform you distribute to customers -- this is for you as the platform provider, not your end customers' own use of the app. Tabs: Dashboard, Products, Licenses, Customers, Activations, API Keys, Provider Console, Events, Audit Log, and (owner-only, on the platform owner's install) Platform Licenses. Platform Licenses has **Licenses** (issue, download, revoke), **Customers** (contracts, invoice drafts, payments, renewals, license updates and usage reports for the customers you license directly) and **Signing key** (rotate the key licenses are signed with; customers need no action, but rotate only after they have upgraded to v4.89.0 or later). **Provider Console** shows a volume tier charge preview for your managed installs and an owner-only form to record each install's usage report, with its history. **Fleet** *(since v4.94.0, owner account on the platform owner install)*: **Route:** `/fleet`. Shows the installations you enrolled for your customers, as they report themselves: version, health, security, AI provider status, licence and last contact, next to your own billing records. Create an installation to get a one-time enrolment string for the customer owner, send plain-text notices that installations can answer, assign informational service packages, and generate service reports. Nothing in Fleet changes a customer's server, and a report e-mail is only sent after you approve it in Approvals. On the **Provider staff** tab you can invite your team into one of five roles (NOC, support engineer, security operator, billing operator, auditor), choose the customers each person may see, and deactivate a person; staff sign in with their own account and see the Fleet page only, with just the controls their role allows. See `docs/UI_GUIDE.md` for the screen reference. --- # 10. Cloud Ops Infrastructure and operations management for your hosted environments. All four screens connect to real infrastructure when credentials are configured; without them, pages show an honest "not connected" state rather than fabricated data. ## MSP Operations **Route:** `/msp` For managed service providers, IT consultants, and anyone managing multiple customer environments. An **Overview** tab shows portfolio KPIs and a health-distribution chart; **Customers** shows per-customer health cards (score, SLA %, ticket counts, ARR, renewal-risk badge) with full add/edit/delete; **Renewals** surfaces high/critical renewal risk sorted by severity, with a **Draft Renewal Outreach** action producing a ready-to-send message grounded in that customer's real account data; **SLA Reports** generates a markdown compliance report per customer or across all; **PostgreSQL** is a monitoring view of your database instances (slow queries, bloat, replication lag). QBR briefings, SLA reports, and renewal outreach are all written by a real AI pass over the customer's actual account data, and account-level data isolation keeps each customer's health/SLA/renewal data scoped strictly to their own account. Attention Feed signals fire automatically for renewals due, SLA at risk, and critical health. ## VMware VCD Operations **Route:** `/vmware` VMware Cloud Director management: full org/VDC/vApp/VM/network/storage hierarchy with CRUD, plus per-tenant billing. **Deleting an organization, vApp, or network is approval-gated** -- clicking delete creates a pending approval rather than executing immediately; the resource stays visible until you approve it on the Approvals page, and the worker executes within a few minutes of approval. ## OpenStack Operations **Route:** `/openstack` Instance lifecycle (start/stop/delete), per-host hypervisor utilization, and per-project cost attribution. Idle instances (14+ days) show a badge. **Deleting an instance is approval-gated** the same way as VMware above -- confirming does not terminate it immediately; it creates a pending approval, and the worker executes the delete after you approve. ## Infrastructure **Route:** `/infrastructure` General infrastructure monitoring -- server health, container status, resource utilization, deployment history -- a unified view across whichever providers you've connected. --- # 11. Governance Controls, security, and platform oversight. ## System Health **Route:** `/system-health` Three tabs. **Live Health** is a real-time dashboard covering every major component -- API, database, Redis, worker, connectors, license validity, security service, sandbox executor, document renderer, and Support channels -- each with a green/amber/red status. **Readiness** (v4.65.0, owner-only to run) answers a deeper question than a config checklist: is this installation actually ready to operate. Each of 8 areas -- core platform, AI providers, the Executive Board, execution/approvals, capabilities, notifications, security, backup -- is checked for real and stated as VERIFIED, CONFIGURED BUT NOT VERIFIED, NOT CONFIGURED, FAILED, or NOT APPLICABLE, never a bare pass/fail. Click **Run Check Now**, or it runs automatically after install and after a real upgrade. **Diagnostics** is the setup checklist and preflight verification: confirms required environment variables are configured, connections are reachable, and migrations are current. Check Live Health daily (or whenever something feels slow); use Readiness before trusting the system with real missions; use Diagnostics during initial setup or after a major configuration change. ## Custom Departments *(Formerly "Departments" -- renamed v4.46.0 to distinguish it from the fixed workforce department pages elsewhere in the sidebar, such as Engineering Department under Workspace -- this screen manages dynamic cost centers, not those.)* **Route:** `/departments` The real department hierarchy -- each department's agent count, trailing-30-day AI spend, and monthly budget with a utilization bar (green/amber/red by how close to budget it is). Create new cost centers here and set their monthly budget. **Full management controls (v4.46.0):** a department card now expands into a real edit view -- rename, adjust budget and alert threshold, delete with confirmation, and manage which agents belong to it, including setting a real department head with a star toggle. A name-collision guard steers you away from naming a cost center after one of the fixed department pages (Engineering, Sales, Growth, Product, Finance). Use this to see which department is driving AI spend, or to set/adjust a budget before it goes over. ## AI Service Desk **Route:** `/incidents` Incident and service-request tracking with AI-assisted analysis. Severity tiers run P1 (Critical, 4h SLA) through P4 (Low, 72h SLA). Open an incident and click **Analyze** to get a root-cause analysis with a confidence score and a remediation runbook, whose medium/high-risk steps each carry an **Approve** button -- approving one is a real governance action that creates an Action, an Approval, and an Audit Log entry, so every risky remediation step stays fully traceable. **Mark Resolved** persists resolution notes to the database. A **Close Incident** button appears once an incident is Resolved -- a distinct, final step beyond Resolved. P1 incidents fire automatic attention signals on creation and on SLA breach (the highest urgency in the system). ## Support Analytics **Route:** `/support-analytics` **(v4.14.0)** Customer Operations dashboard for support cases -- case volume, open backlog, SLA breaches, escalation rate, per-skill performance, and blended cost-per-resolution, filterable by venture. Check weekly for trends, or when verifying a specific support skill's performance. ## Backup Center **Route:** `/backups` Backup management -- schedule, status, and restore for database snapshots and configuration exports. A daily full backup runs automatically (database dump, vault archive, config archive, combined with a SHA-256 checksum, stored locally with optional S3/NFS upload); create additional schedules with a cron preset or a custom expression, choosing backup type (full/db/vault/config) and retention period. On any completed job: **Verify** re-checks the checksum, and **Restore** opens an inline confirmation with a **Validate (dry-run)** option that checks archive integrity without touching live data, plus a full **Restore Now**. A backup with no schedule (from **Backup Now**, or from a schedule you deleted) is kept while it is among the newest 10 or younger than 30 days, then deleted automatically with its archive file; the newest completed backup is never deleted automatically, and the Jobs tab shows how many backups the next daily run will delete. The owner can also delete a completed backup with **DELETE**: the archive file is deleted too and the backup can no longer be restored. Verify weekly that backups are completing, and test a restore quarterly. ## Audit Log **Route:** `/audit` Two tabs. **Full Audit Log** is the immutable, append-only record of every system action -- who did what, when, with what result; it cannot be edited or deleted. **Actions Journal** is a filtered view of agent-initiated actions specifically, easier to scan day-to-day. Use the full log for compliance and security review, and the journal for routine agent-activity review. --- # 12. Workspace The operational workspace for projects, engineering, growth, and internal process automation. ## Conversations **Route:** `/conversations` Two tabs. **History** logs every conversation between you and agents, searchable by agent, date, or keyword. **Agent Chat** is a live chat interface -- select any agent and send it a direct task or question; the response streams and genuinely runs as the selected agent, not always the CEO. It has a mic button for voice input and a paperclip for document attachment, and any pending approval for that agent shows inline in the conversation instead of requiring a trip to the Approvals page. When an agent generates a document or presentation here, the reply includes a real, clickable download link. You can also ask for a pure business goal here or from the command bar (see Overview, "Issuing Commands") and get a ranked shortlist of real opportunities back. ## Projects **Deploy waits for approval (v4.91.0):** choosing **Deploy** on a project creates a pending approval in Approvals; the build starts only after an owner approves it, and the page tells you that approval is needed. **Route:** `/projects` Deleting a project is owner-only. The confirmation states that the project, its containers and images, and its source folder are removed and that this cannot be undone; if another project of your workspace uses the same folder, the files are kept while it exists. Real, running software projects ErezAgent builds and operates for you, each backed by its own isolated sandbox container. Deploy, stop, and monitor build logs and health from the list; a project moves through scaffolded → building → running (or failed, with an auto-restart retry loop before it's marked failed and an urgent attention signal fires). If a build genuinely crashes partway through, the mission's own conversation now narrates the failure plainly so you know it needs a retry, instead of the run just going quiet. Each project's detail panel has four core tabs -- **Overview** (status, health, version, ports, internal URL, originating mission and real cost-to-build, linked deal value when tied to one, a **Publish** panel covering automated Cloudflare/DNS publishing, manual/BYO publishing, or an Unpublish control once live, and a **Request Change** panel that describes a change in plain language and re-enters the build/test/fix pipeline against this exact project), **Architecture** (the deterministic 12-field stack record an agent loaded before writing any code), **Decisions** (every significant technology choice logged, with type, impact level, and rejected alternatives), and **Timeline** (a chronological journal of deploy/build/decision/revenue/customer/incident events). The **Customer Asset** tab is where a customer-solution-delivery project gets deployed and formally accepted -- see Mission Control (Group 3) above for the full requirements-to-blueprint-to-acceptance flow, including the Android delivery path. ## Engineering Department **Route:** `/engineering` Every launched venture automatically owns a persistent Engineering Department; select a project from the dropdown at the top to view or initialize one manually for an existing project. Six tabs. **Department** shows the engineering agent role cards (Engineering Manager, Software Architect, Backend/Frontend/DevOps/Database/Security/QA/Performance engineers, Release Manager, Technical Writer) plus a compact Quality Overview and a **Repository Health** widget -- a composite 0-100 score from test coverage, dependency security, code complexity, and open TODOs, with a delta against the previous scan, file/line counts, vulnerability count, and a language breakdown; click **Rescan** for an immediate fresh scan (it also runs automatically every few hours). A detected issue auto-creates a matching Tech Debt item (a vulnerable dependency as critical/security, low coverage as high/testing, high complexity or excess TODOs as medium/low code-quality items). **Sprint** is a Kanban board (To Do / In Progress / In Review / Done); a **Parallel Work Streams** panel at the bottom lets the Engineering Manager break a large task description into parallel streams across four dependency waves (database first, then backend/auth, then frontend/tests/docs/deployment, then monitoring) -- type a task and click **Decompose**. **Backlog** lists unassigned work items. **Tech Debt** is the debt register with severity counts and per-item business/security impact. **Quality** shows 8 score cards (Architecture, Maintainability, Security, Performance, Testing, Documentation, Deployment Health, Repo Health) plus operational metrics (velocity, lead time, cycle time, bug escape rate). A **Reviews** tab shows the automated review chain that runs whenever a code change is ready to ship -- five reviewers in sequence (Architect checks repo health, Security checks vulnerable dependencies, QA checks test ratio, Performance checks complexity, Release Manager checks sprint progress), each returning approved/concern/blocked with a confidence score and specific findings. A session is **Approved** (all hard-block checks passed), **Blocked** (a hard block is open, with the specific issue shown), or **Bypassed** (a recorded, permanently auditable override reason). Click **Run Review** to trigger the chain on demand. An **Evolution** tab watches your project's engineering health and proposes specific improvements -- click **Run Analysis** to read work-stream history, repo health, and review sessions, and get back prioritized proposals (Security Hardening, Process Change, Architecture Change, Roadmap Item, Refactor, Performance) you can **Approve**, **Add to Sprint**, or **Reject** with a reason. ## Growth Department **Route:** `/growth` The Growth Division dashboard, scoped to one venture at a time. **Overview** shows a Growth Health Score (0-100) composed of 8 gauges -- SEO, Brand, Traffic, Content, Social, Ads, Email, Web Performance -- and the full department roster (18 roles, seeded automatically per venture). **Campaigns** are time-boxed initiatives with a budget, channels, and a goal, moving through draft → active → paused; the spend/budget bar turns orange past 70% and red past 90%. **AI Draft** on a campaign card opens a modal to generate a real content draft (social post, SEO/GEO content, or an email sequence) grounded in that campaign's own goal and audience. **Creative** is a 5-column Kanban board (Backlog → In Progress → Review → Approved → Published) for content tasks. Any generatable task type -- blog post, social post, email, ad creative, landing page, whitepaper, PR outreach -- can be AI-generated with one click; publishing is a high-risk action that always routes to Approvals before a task can move to `published`. A creative task backed by an AI-generated image shows a real thumbnail and a critique-quality badge, with a **Regenerate** button for another version at any time; a coordinated multi-asset brief groups its cards together in the board. **SEO** shows your live SEO debt backlog (severity, category, description, an AI-generated fix proposal) and a health-score sparkline; **Run SEO Scan** triggers an immediate check (also scheduled daily) that calls PageSpeed Insights, crawls your homepage for missing metadata and broken links, and recomputes the health score. **Design Studio** is your brand asset registry -- logos, colors, fonts, templates, images, icons, illustrations, video, and short-form brand copy -- each starting `pending` until you explicitly approve it, with version lineage via a parent-asset link and an optional Obsidian path reference. Since v4.36.0, a customer you're delivering software for can have its own separate brand identity here too, reachable from that customer's own page in chat rather than a venture's Growth Department page -- and a delivery project for that customer automatically pulls its real colors and assets into the design work instead of inventing a look from scratch. **Intelligence** shows real, reconciled results from Growth's own autonomous demand-generation loop -- spend, revenue, and ROI from experiments the CRO/Growth agents have actually run and measured, kept distinct from manually-entered campaign metrics and staying honestly empty until a real experiment has been measured. A KPI row (total spend, total revenue, average ROAS, average CAC), signal cards (an active campaign with spend but no attributed revenue, or ROAS under 1.0), a per-campaign ROAS bar chart, and a health-score trend line round it out. The CRO agent can also qualify and quote a lead on its own as part of this loop, and an optional Meta Ads connector lets Growth create real paid campaigns, always subject to your approval and the venture's spend limits. ## Social Inbox **Route:** `/social-inbox` Every inbound social conversation -- public comments, mentions, and (for Facebook, Instagram, and Twitter/X) direct messages -- in one list-and-detail view. The list shows status, channel type, and platform; the detail pane shows the full thread, what it was routed to (a real lead, support case, or product signal), and a Close/Reopen action with no approval gate, the same class of direct action as resolving a support case. Review daily to see what customers and prospects are actually saying without hunting through each platform separately. ## Product Department **Route:** `/product` The product-management operating loop: **Signals** (support cases, Growth findings, Opportunity Radar scans), **Opportunities** (clustered from signals and deterministically prioritized), **Roadmap**, and **Learnings**. **Recompute Priority** on the Roadmap tab re-scores and re-ranks every open opportunity with a RICE-style formula; **Evaluate Adoption** on a released opportunity runs a real post-release adoption assessment; **Generate Lesson** on the Learnings tab writes a lesson from a completed opportunity's outcome, feeding future discovery the same way Strategic Lessons does elsewhere. Commitment always goes through a real Executive Board review with a genuine CPO position -- use this page when deciding what to build next and tracking whether what shipped actually worked. ## Company DNA **Route:** `/company-dna` The accumulated organizational knowledge graph -- patterns, decisions, culture, and operational practices extracted from interactions over time, growing automatically as the organization operates. Use it to understand the organization's institutional memory: what it's learned, and how it actually operates versus how it's documented to operate. ## Workflows **Route:** `/workflows` Two tabs. **Workflows** lists your workflows (Active, Errored, Inactive) with a venture filter; click a workflow to see the history of its runs -- status, step completion, timing, cost -- and use its details button for **Versions**, **Steps**, **Triggers** and **Settings**. **AI Planner** turns a plain-English objective into a step-by-step execution plan (a DAG) for review and approval before it executes. A workflow can run steps internally (light automation) or trigger an n8n webhook for visual/external automation. Workflows can also fire automatically on real system events (a new approval, an approval accepted, a new mission, another workflow completing, a new security finding, and -- since v4.81.0 -- a business record being created, changing stage, or matching an automation rule), with optional filter conditions restricting when a trigger fires. A workflow step can now also genuinely pause (v4.81.0): a "wait" step holds for a set duration and resumes exactly where it left off, rather than every step having to run start to finish in one go. Every workflow carries a retry policy (max attempts, fixed or exponential backoff) and can declare a rollback workflow to run automatically if every attempt fails; an SLA in seconds triggers an Attention Feed alert if a running workflow exceeds it. A workflow can link to a mission or venture goal so the Outcome Engine (Group 7) can attribute its runs to a real business outcome. **Versions, steps and approval (v4.88.0):** every save of an internal workflow's steps creates a new version. The **Versions** tab lists them and an owner can **Restore** an earlier one; restoring writes a new version that copies the old one, so history is never rewritten. Besides the existing steps, a workflow step can now create a task, send a message (email or WhatsApp), add a document requirement, invoke a skill, request a signature, or wait. Steps are checked when you save or activate: an unknown step type is refused, and a run that meets one stops rather than continuing. A workflow can be tied to one venture (it then acts only on that venture's records), and a business-record trigger can be limited to one record type. A workflow marked as requiring approval now genuinely waits for approval in the Approvals queue before it runs; the **Requires owner approval to run** checkbox in the workflow's Settings can be changed by the owner only. Deleting a workflow that is still referenced asks for confirmation and is owner-only. ## Business Records **Route:** `/business-records` Define your own configurable record type for anything that doesn't fit CRM, Support, or Appointments -- vendor onboarding, an internal review process, whatever your business needs to track. Give it a name, add custom fields (text, number, date, checkbox, single-select, multi-select, file, or signature), and define a stage workflow: which stages exist, which one records start in, which stages are final, and which moves between stages are allowed. The left column lists your record types; click "+" to open the type builder. The middle column lists records of the selected type; click "+" to create one, filling in the fields you defined. Selecting a record opens its detail view, where you can move it to any stage your workflow allows from its current stage, with an optional reason, and see its full stage history. Editing a type later creates a new version rather than changing existing records -- a record already created keeps validating against the exact version of the type it was created under, so adding a new required field to a type never invalidates records made before that change. **Rules (v4.81.0):** click the lightning-bolt icon next to a selected record type to open its Rules panel. A rule has one or more conditions (checking a field's value or the record's current stage), combined with AND or OR, and one or more actions: move the record to another stage, hand it off to a teammate as a real to-do, (v4.82.0) request a signature, and (v4.88.0) **create a task** (for a person or a role, with a due date and priority), **send a message** by email or WhatsApp, **add a document requirement**, or **invoke a skill**. Messages and signature requests always wait for approval before anything is sent, and a message is not re-sent to the same record within its repeat window. When a matching record is created or moves stage, every active rule for that type is checked automatically, and every match is logged in the panel's activity history. **Time-based rules (v4.88.0):** a condition can also use **days in stage**, **days since update** or **days since created** (numbers only), for example "in Review for more than 5 days". These rules are checked every 30 minutes, are marked "time-based" in the panel, and fire once per time a record enters a stage, so a record that stays put is not acted on repeatedly. When you save a new version of a record type, its rules and customer portal pages are carried forward to the new version; a rule that no longer fits the new fields or stages is copied as a draft for you to fix. In the type builder, each field also has a "visible to" control -- restrict a field to specific roles and everyone else simply won't see or be able to edit it. You can also ask Executive Office directly: "how many open vendor onboarding records" or "what has automation done today," by chat or voice. **Documents and e-signature (v4.82.0):** a record with a file field lets you attach a real document; a record with a signature field lets you send one of those documents out for signature (owner/operator only -- every signature request goes through the normal approval flow before it's sent), and its status updates automatically as the signer acts (sent, delivered, signed, or declined). After attaching a file, use "Suggest values" to have AI propose values for the record's other fields based on what's in the document -- nothing is applied until you accept it, field by field. If two documents on the same record disagree on a value, or a document is approaching or past its expiry date, it shows up in the Documents panel for you to review and acknowledge. Ask Executive Office "what documents need attention" for a portfolio-wide view, or ask about a specific record's signature status directly. **AI-drafted types and templates (v4.83.0):** don't want to build a record type field by field? Click "Ask AI to build a type," describe what you want to track in plain language, and AI drafts a complete type (fields and stage workflow) for you -- it's sent for your review and approval before anything is created, the same way every AI-drafted change in ErezAgent works. Or click "Browse templates" to install a ready-made type (Vendor Onboarding, Client/Matter Intake, Contract Review) instantly, no review needed since these are pre-built by ErezAgent, not AI-generated on the spot. You can also pin a specific AI capability to a type's own document-extraction calls in the type builder, if the default isn't giving you the quality you need for that particular kind of document. **Downloading files (v4.87.0):** a file field on a record has a **Download** button that saves the original document; each download is recorded in the Audit Log. **Sectioned workspace (v4.85.0):** open any record to see it laid out in sections -- its fields, documents, tasks, appointments, approvals, and full stage history together on one page, instead of hunting across other pages for related work. Bookmark or share the direct link to a specific record (`/business-records?id=...`) and it opens straight to that record. **Customer Portal (v4.85.0):** give a record type its own customer-facing portal -- select the type, click the globe icon next to Rules to open its **Customer Portal** panel, and add which sections a customer can see and act on (their record's fields, a document upload, appointment booking, and three read-only status views: a task checklist, signature status and approval status). The task checklist and approval status show only tasks you tick **Visible to customer** on in the record's Workspace; signature status shows only signature fields tagged `customer`-visible. A customer reaches their own record at a link scoped to their venture and that one record only -- they can never see or reach any other record, even another one of yours. Only a field you've marked `customer`-visible in the "visible to" field control above can ever appear in the Record fields component -- an untagged field is never shown to a customer, even if it's visible to every internal role. **More portal components (v4.88.0):** **Stage progress** (where the record is in its stages, with no internal notes), **Required documents** (what is still needed, with upload through the existing document page), **Customer form** (a form for the fields you mark **customer editable**, only in the stages you choose; a customer cannot overwrite a newer change made by your team, and each edit is recorded), **Messages** (the customer's conversation with you about this record) and **Requests** (the customer can raise a request, which becomes a support case linked to the record). Like the others, each is available only for that customer's own record. ## My Work **Route:** `/my-work` *(Since v4.85.0.)* One itemized list of everything on your plate -- your open tasks, pending approvals, and open loops -- pulled from across the system into a single view instead of checking each page separately. Filter by status or item type; each item links straight back to where it lives (its task, approval, or mission step). If you manage staff across one or more ventures, a team workload panel shows who's carrying the most open work. ## Operational Dashboard **Route:** `/operational-dashboard` *(Since v4.85.0.)* A read-only operational health view across your (or your staffed venture's) business records: which records have gone stalled (no stage movement for a while), how work is distributed across your team, and where records tend to bottleneck by type and stage. Use it to spot where a process is backing up before it becomes a customer-facing problem. ## Goal Planner **Route:** `/goal-planner` AI-powered goal decomposition for multi-stage business goals. Type an ambitious goal and Claude analyzes it into 2-5 ordered stages, each its own full execution plan; **Approve & Execute** runs them sequentially -- stage 2 only starts once stage 1 completes, and a failed stage skips its dependents unless configured otherwise. Use this when a goal is too large for a single workflow and genuinely needs sequential phases; for a single objective that can run its steps in parallel, use the Workflows AI Planner above instead. --- # 13. Platform System configuration, user management, and platform administration. ## Platform Management **Route:** `/platform` The platform administration center, covering license status and system vitals, your license tier and comparison, resource usage vs. limits, a license-events log, Diagnostics, Localization (company locale, market profiles with RTL/currency/timezone), Deployment (deployment model, release channels, installer repository, upgrade workflow), What's New, Migrations, an Update Planner, a Support Bundle export, a Benchmark, Recommendations, Installer info, Billing, and a Mail Log. Use it for initial setup and ongoing administration -- Licensing for your current status, Usage for approaching limits, Mail Log to confirm email delivery. ## Connectors **Route:** `/connectors` Connect and manage every external account ErezAgent can use -- email, calendar, CRM, payments, and social platforms among them. Each connector card shows real connection status, what it's actually authorized to do, and a manual sync/reconnect option. All credentials are encrypted at rest. **Connecting a service:** click Connect on the card, enter the specific credential it asks for (a bot token, a secret key, and so on), and the card moves to Online. Gmail and other OAuth connectors redirect you through the provider's own authorization page instead. If you manage multiple ventures, a **Scope** dropdown lets you give one venture its own separate credential (a second Gmail inbox, a different Stripe account) without disturbing the shared connection every other venture still uses -- works for both API-key and OAuth connectors. **Status meanings:** Online (connected, health check passed), Auth Required (credentials exist but authentication failed -- reconnect), Disconnected (no credentials saved), Error (health check failed, service unreachable). **Social media:** Facebook, Instagram, Twitter/X, TikTok, YouTube, and LinkedIn do more than publish -- once connected, ErezAgent reads inbound comments, mentions, and replies automatically, classifies them, and routes real leads, support questions, and product feedback to the right place (Social Inbox, above), with no separate setup step beyond connecting the account. Facebook, Instagram, and Twitter/X can also read and reply to direct messages -- an already-connected account needs reconnecting once, since the message-read permission isn't granted retroactively. Drafted replies always go through Approvals before anything posts. Connections that could expire (Gmail, YouTube, Twitter, TikTok, DocuSign, Google Analytics, and others) are checked and refreshed automatically before they lapse. Use this page to connect a new account, spot a stale connection (look for `Auth Required`), or reconnect one that's expired. If a card shows "Missing permission scope", the account was connected without a permission ErezAgent needs (for YouTube, managing comments); background reading of that account pauses until you click Reconnect and grant access. A connection made for one venture is named in the note and reconnected on its own. Only owners can reconnect. **Builder tab** *(since v4.50.0):* need to connect to a system that isn't already on this page? Define it yourself -- its address, how it authenticates, and the specific actions it supports -- without waiting on an engineering release. A new connector starts as a draft and won't be used for anything real until you explicitly activate it; **Test** runs a real connectivity check against it at any time. ## Operational Workspaces **Route:** `/workspace` *(Since v4.50.0.)* Where you monitor and act on an ongoing operational commitment ErezAgent manages on your behalf -- a pricing floor, an SLA credit policy, and similar rule-driven missions that keep running after a project delivers, not just a one-time build. Each workspace shows its live queue, full decision history, and anything genuinely needing your review (a rejected recommendation, a low-confidence match). A new workspace appears automatically when a delivered solution needs one. ## AI Resources **Route:** `/resources` The AI Resource Manager -- registers and monitors every AI provider (Anthropic, OpenAI, Google, Ollama, DeepSeek, xAI Grok, Mistral, and more) as a resource with capability scores, availability, and cost data. Instead of hardcoding a model, an agent declares what it needs ("reasoning and coding, under $5 per 1M tokens, must stay private") and the router finds the best match. **My Resources** shows every registered provider with live availability, execution mode (API / Local / Non-LLM API), cost tier, a Privacy Safe badge for local providers, and per-capability quality scores; click **Test** to probe a provider immediately. **Capability Map** is a full grid across reasoning, coding, vision, research, speech-to-text, text-to-speech, embeddings, long context, local, fast, structured output, market data, software engineering, frontend engineering, computer use, and browser interaction -- each cell showing which providers cover it, plus a **Test a resolve()** panel that shows exactly which resource would be chosen for a capability and why, using the router's real scoring breakdown, never a fabricated explanation. **Discovery** lists models a provider's catalog mentioned or that have a configured key but no verified real call yet -- never auto-selected for real work until you click **Promote to Active**. **Effectiveness** shows real 30-day outcomes per (resource, capability): tasks attempted, pass rate, escalation rate, cost, and a verified value-per-dollar figure that honestly reads `INSUFFICIENT_EVIDENCE` rather than guessing when there isn't enough real data yet. **Usage & Cost** rounds out 30-day analytics: total requests, total cost, fallback count, and a breakdown by capability and resource. If you're processing sensitive data, any agent can request `private_data: true` and the router will only select a provider marked `privacy_safe` -- in practice, a locally-run Ollama instance, so your data never leaves your machine. **Subscriptions and MCP servers:** interactive subscriptions (Claude Max/Pro, ChatGPT Plus, Gemini Advanced) can't be auto-called in pipelines, but registering one tracks its monthly cost and capabilities for visibility, and shows it (grey, non-routable) on the Capability Map. A **Manual Handoff** flow bridges the gap: package a prompt, copy it into the subscription's own interface, paste the result back -- useful for a subscription-only capability like a large context window you want coordinated through ErezAgent anyway. MCP servers register by URL with an immediate connectivity test; once connected, their tools register as real capabilities. **Discover** scans your environment for installed tools (Claude Code CLI, Ollama, Node.js, Python) read-only, to help you decide what's worth registering. Each resource card shows a live health indicator (Healthy, Stale, Degraded, Unknown), rechecked every few minutes; when a primary resource goes down, a degraded-resource event appears in your event stream. Any agent with declared required capabilities is automatically routed to the best available resource at execution time, falling back to its default model with a warning if nothing currently covers what it needs. **Provider spend/balance (since v4.77.0, card added v4.86.0):** beyond ErezAgent's own internal Usage & Cost analytics above, a **Provider Spend/Balance** card on this same tab shows the real number your AI provider itself reports -- how much you've actually spent with Anthropic or OpenAI over the last 30 days, or your real remaining prepaid balance with xAI -- via a separate Admin/Management API key for that provider (never the ordinary inference key you already entered; see `env.example`'s `ANTHROPIC_ADMIN_API_KEY`/`OPENAI_ADMIN_API_KEY`/`XAI_MANAGEMENT_API_KEY`/`XAI_TEAM_ID`). Each row is honestly labeled **balance**, **spend estimate**, or **unavailable** -- an estimate is never shown as if it were a real remaining balance. Gemini has no such API, so it falls back to ErezAgent's own internal estimate instead, clearly labeled as such. A row also discloses its real scope: **install**-wide for the three external-API lookups (a shared Admin key reports for the whole install, not per tenant, on a multi-tenant deployment), or **tenant**-scoped for Gemini's own internal estimate. ## Workforce AI **Route:** `/workforce` AI workforce analytics -- agent cost allocation, utilization rates, and performance benchmarks across the full agent workforce, with a 7D/30D toggle. A tier leaderboard (Elite, High, Underperforming, Probation, Suspended) ranks agents by ROI and accuracy; an Agent Cost vs. Value table shows cost, attributed revenue, ROI ratio, waste percentage, and a recommendation per agent (expand authority, high value, review allocation, consider suspend). Workforce Actions propose model promotions/downgrades, suspensions, and budget adjustments for your Approve/Reject. Review monthly to spot underutilized agents or disproportionate cost. ## Docs **Route:** `/docs` The built-in documentation viewer -- renders every design doc, guide, and changelog from the system's docs directory directly in the app, across multiple tabs including the EULA and Privacy Policy, with a live search bar that matches whole sections (H1/H2/H3) by keyword. Use it to reference documentation without leaving the application. ## Users **Confirm with your password (v4.91.0):** creating a user asks you to type your own password in **Your password (to confirm)**. A workspace always keeps at least one active owner. **Route:** `/users` User management -- owner role only. Add, deactivate, and reset passwords for team members; assign owner/operator/observer roles. See Roles and Access in the Overview section above. ## Venture Staff (v4.84.0) Separate from Users above: each venture has its own **Staff** panel (open a venture on the Ventures page and click **Staff**) for inviting operational employees who only need to work that one business, not your whole account. Invite by email as a Staff Member or Staff Manager; they set their own password from the invite link and sign in through the normal login page. A staff account can read and write that one venture's business records, review and act on customer-submitted documents, manage support cases, and view/manage appointments and tasks -- and nothing from any other venture, even one you also own. Revoke access from the same panel; it takes effect on their very next request. Staff seat limits apply: a total across all your ventures, plus an optional cap on any one venture if you want to divide seats unevenly. **Staff roles (v4.88.0):** open a venture's **Staff** panel and switch to the **Roles** tab to define named roles for that venture (for example "Case handler" or "Reviewer"). For each record type, a role can be allowed to view, create and edit records, move them only to certain stages, and see all records or only those assigned to them; a role also lists the pages and actions (such as request signature, change stage, assign task, review documents) its holders can use. On the **Team** tab, choose a role when you invite someone (the invite cannot be sent until you pick one, or pick **No role (full staff access)** on purpose) or change it later; **No role (full staff access)** keeps the original behavior, so staff invited before this release are unaffected. Editing a role saves a new version; archiving one asks you first, and members who held it have no access until you give them another role. Role changes apply on the staff member's very next request. Defining roles is owner-only. A field restricted by role in a record type stays hidden from roles that are not allowed to see it, and staff can also open the venture's operational workspace views you have built. **Where staff actually work (v4.85.0):** a staff member signs in to the same My Work and Business Records pages described above, automatically scoped to only the venture(s) they're staffed on -- they never see a venture selector for a venture they aren't a member of. A **Staff Manager** additionally sees their venture's team workload on the Operational Dashboard. **Staff workspace (v4.86.0):** after signing in, a staff member lands in a focused workspace showing only My Work, the Operational Dashboard, Business Records, Document Review, Incidents, and Appointments, and signs out from the same workspace. Customers can also check their document checklist and upload files from a link in their business-record portal. **Staff Incidents and Appointments (v4.87.0):** staff see the customers and ventures they are allowed to use in these pages' pickers, owner-only appointment setup controls are hidden, and staff can run AI outcome extraction on appointments in their own ventures (the result is a draft that someone still confirms). Revoking a staff member asks you to confirm first. ## Settings **Route:** `/settings` User preferences: notification settings, display options, API key configuration, session management, budget and security-gate thresholds, and a Skill Creator Auto-Trigger toggle (off by default) that turns on a daily job proposing skill revisions on its own -- flips live, no restart needed. A **Voice Interaction** section covers voice output, TTS provider and speed, Executive Conversation Mode's barge-in toggle, its silence threshold, and a Proactive Notifications toggle (off by default). **Voice Profiles** lets you clone your own voice or pick a curated preset for ErezAgent to speak with, running on your own hardware via a free, open-source local voice server you run yourself -- there's no bundled/managed version; point the relevant setting at your own instance. Cloning needs a short sample plus the exact reference text spoken in it; a preset just needs an engine and voice choice. One profile is active at a time -- that's the one ErezAgent speaks with when Voicebox is your selected TTS provider. Since Persona Management (Group 6) shipped, a Persona can also be given its own distinct voice pulled from this same profile list, independent of your account-wide default. **OpenAI voice:** with OpenAI selected, speech-to-text uses `gpt-transcribe` and spoken replies use `gpt-realtime-2.1-mini`; if a model is not available to your key, the previous one is tried once. Spoken replies start playing while they are still being generated. ElevenLabs and Voicebox are unchanged. Two-Factor Authentication setup lives here too -- see Roles and Access in the Overview section. **Data & Privacy (owner only):** the **AI analysis of your documents** card turns AI reading of the documents you upload on or off (off by default) and chooses the mode -- Redacted (identifiers masked before text is sent to your AI provider) or Local models only. See Build My Business (Group 2) for what it covers. The same section holds the contact export and erase tools: erasing a contact now covers every table that holds personal data, and can match on a phone number as well as an email. **License (v4.89.0):** the Platform License section shows the installed license's edition, status, expiry and limits, or your trial and its days remaining. A licensed install never shows the trial screen. To install a purchased license, the owner chooses **Enter License Key**, pastes the license file and presses **Activate**; a file that is unsigned, edited or expired is refused with a message and nothing changes. **Trial and license reminders (v4.91.1):** a trial banner appears from 14 days before the trial ends, and an installed license's banner appears from 30 days before it ends; the banner shows the days remaining with **Renew / Buy license** (opens the vendor's renewal page), **Request license by email** (opens an email to the vendor with your machine fingerprint, the end date and your installation owner's name and email already filled in) and **Enter license**. It is amber at first and turns red at 3 days. You can dismiss it for the rest of the day; it returns the next day, and at 1 day left or on the last day it cannot be dismissed. When a trial or license has ended, a "Trial Expired" or "License Expired" screen explains in three steps how to continue: get a license, paste it, and select **Activate License**. Your data is kept untouched and everything returns as soon as a license is entered. The machine fingerprint is shown there with a **Copy** button, and only the owner can paste a license. **Provider / Managed Service (v4.94.0, owner only):** if a provider manages ErezAgent for you, this section connects the installation to it. Paste the enrolment string the provider gave you and select **Check enrolment string**; the section shows the address that would receive the report. Read the lists of what is shared and what is never shared, tick the consent box and select **Connect**. Nothing is sent before then. If a later version changes the list, reporting pauses until you confirm the updated list. Afterwards the section shows the provider's name, the managed-install ID, the last report, the support contact and the latest version the provider knows, with **Show what is reported** (the exact report last sent), **Pause reporting**, **Resume reporting** and **Disconnect**. Notices from the provider appear here and in a strip at the top of the page, with four answers: I will update, Remind me later, Request assistance, Request professional services. The provider can see operational state only and cannot change anything on your server. See ADMIN_GUIDE.md section 24. **Plan & Billing (v4.51.0, updated v4.89.0):** this install's plan, limits and billing, with a badge for what governs it: **Platform owner instance - everything enabled** (every limit unlimited), **Licensed install** (plan, status and expiry read from the license; capacity and add-ons are changed by your vendor through an updated license), **Subscription** (billing term, upgrade or downgrade with proration shown first, add-ons that stay pending until payment is recorded, promo codes, cancel at period end) or **Trial**. Only an owner sees the controls; see ADMIN_GUIDE.md section 11 for billing methods, Billing Setup and invoicing. --- # Troubleshooting | Problem | Likely cause | Fix | |---|---|---| | System Health shows the AI Provider in mock mode | Running without a live model provider | Set `ANTHROPIC_API_KEY` and leave `MODEL_PROVIDER_OVERRIDE` empty in `.env`, then rebuild | | Agent responses are generic, not specific | Mock provider active | Same as above | | Login fails | Wrong password, or migrations haven't run | Check `ADMIN_PASSWORD` in `.env`; run migrations | | Live-update indicator is offline | Redis down or connection dropped | `docker compose restart redis api` | | Approval badge not updating | Live connection interrupted | Check the connection indicator; refresh if needed | | Telegram notifications not arriving | Wrong bot token or chat ID | Reconnect Telegram in Connectors; verify the owner chat ID | | Obsidian sync fails | Wrong URL inside Docker | Point `OBSIDIAN_API_URL` at your host, not `localhost` | | Outcome Engine shows $0 | No `closed_won` deals yet | Close a deal in Revenue Center; revenue events appear within about 15 minutes | | AI cost seems high | Budget limits not tuned, or an expensive model on a low-complexity task | Review Token & ROI; approve any pending model-downgrade proposal on Self-Improvement | | Agent not responding | Worker container down | `docker compose restart worker` | | Document upload fails | File too large or unsupported format | 50 MB limit; supported: PDF, DOCX, XLSX, PPTX, CSV, MD, TXT, ENEX | | Build My Business: Analyze is disabled | The list under the buttons says why: owner role missing, AI-analysis consent not granted, no files, or cost not yet estimated or confirmed | Turn on consent in Settings > Data & Privacy (owner), add files, then Estimate cost (and confirm it if asked) | | Build My Business: a file shows "Could not be read" or "Skipped" | Unsupported format (.doc, .xls, .ppt, Outlook .msg) or a file with no readable content | Save as .docx/.xlsx/.pptx or export the message as .eml, then upload again; the reason is shown under the file | | Build My Business: analysis stops in local-only mode | No locally hosted model is available (images need a local vision model) | Start your local model server, or switch the mode to Redacted in Settings > Data & Privacy | | A scanned or photographed customer document sits in the Review Queue | AI-analysis consent is off (photos), or the PDF has no text layer | Grant consent to have photos read by AI, or review the file yourself | | Approval expired before you could act | Expiry window too short for your workflow | Adjust `approval_timeout_hours` in Settings; re-trigger the action | | Memory search returns weak results | Too few memories, or embeddings still queued | Upload more source material; wait for the embedding worker | | Connector shows Error after saving | Wrong token, or the service is unreachable | Click Sync to see the live health-check error; verify the token in the provider's own dashboard | | A cloud provider page shows empty/"not connected" | Credentials not set in `.env` | Add the provider's credentials, then `docker compose up -d api` to reload env vars (no rebuild needed) | --- *© 2026 Erez Rozenbaum. All Rights Reserved.*