Team briefing
What we are building, why it is different, where it stands, and what happens next — for anyone joining the project or explaining it to someone else. No engineering background needed.
Provenance note, deliberately visible. This page was authored fresh from the repository — the plan, the execution board, the gap ledger and the code itself. It is not a copy of the earlier "team briefing" artifact: that page lives in a different Claude organisation and cannot be read by the build agents at all, so copying it was impossible. Every claim here traces to a file in the repo. If the original contains anything not reflected below, paste it in and it will be merged.
01The one-paragraph version
We are building an AI colleague for teams whose entire life is one append-only diary. Every thought, action, mood, permission, painted card and delivered file is an event in that diary — so every screen is just a replay of it, and you can rewind the whole colleague to any moment and see exactly what it knew, did, and was allowed to do. That is the wedge against every "AI employee" competitor: not another chatbot, but an employee whose work is audit-grade by construction — provable, replayable, and therefore safe to grant real autonomy.
02Why this is hard to copy
Rewind is not a feature bolted on — it is the architecture
Because nothing is stored except the diary, and every screen is computed from it, "show me this workspace as it was on Tuesday at 2:14pm" costs us almost nothing. A competitor storing current state cannot add this later without rebuilding from scratch.
Permission is evidence, not configuration
Every consent is an immutable event. When a buyer asks "how do we know your AI didn't grant itself that permission?", the answer is a replayable chain, not a policy document. If a consent record is ever corrupted, autonomy freezes shut rather than failing open.
Honesty is machine-enforced
Seven anti-drift laws run in CI. One of them forbids silently swallowing a failure: when something breaks, the colleague records it and says so. We burned 43 such paths to zero and the build now fails if a new one appears.
Files carry a birth certificate
Ask for a chart and the colleague runs real code in a sandbox, delivers the file, and pins a card. Click it and you see the exact chain of events that produced it. A card someone tries to fake reads "unverified".
03Where it stands today
| stage | state | what that means |
|---|---|---|
| Stage 0 — safety | COMPLETE | CI, the seven laws, replay integrity, credential grenades defused, the pack engine. The floor everything else stands on. |
| Stage 1 — the colleague | ALL BUILDABLE ROWS DONE | Vitto Live, the Space wall, living artifacts + certificates, ask-memory with citations, the rewind bar, earned autonomy, the glass dossier, dark mode, goals, groups. |
| Security + permissions | LANDED 07-27 | The chat preview escape closed; cross-workspace reads sealed on every path including rewind; trust-reset governance fixed; a permanent census that catches any screen wired to the wrong permission. |
| Stage 2 — the team | FROZEN BY DESIGN | Real-time multi-user, one-tap publish, guest agents, device fleet. Frozen by the one-wave law until the ten-team demo, so the current wave gets finished properly. |
| Stage 3 — Studio | NOT STARTED | Users and the colleague author new apps and wall elements. Specs deliberately unwritten until Stage 2 exits. |
Health at the last checkpoint: 1,587 unit tests, 295 interface tests, 606 integration tests, 24 graded exams and 105 browser tests — all green, on every merge, enforced by GitHub. 44 known gaps are open and every one has a named owner; 30 are closed. Nothing is "known but unowned" — that is a law, not a habit.
04How we work — and why it keeps catching real bugs
Every significant change is reviewed adversarially before it is written: independent reviewers try to refute the design, and a judge rules. This is not ceremony — it has repeatedly killed designs that looked correct. Recent examples, all real: a fix for cross-workspace data access would have looked sealed while the rewind path still leaked, because rewind reads history a different way. A proposed security tightening would have made years of stored history permanently unreadable. A trust-level rule was written backwards — it would have granted autonomy to automation and denied it to people. Each was caught before a line of code shipped.
The same discipline applies to claims. When a reviewer measured the chat preview vulnerability in a real browser, it turned out our own description was wrong — it does not steal login tokens, it does something different and still serious. We corrected the record rather than keeping the scarier story.
05What a customer actually gets
The Friday board meeting
"Chart our MRR by month and pin it." When an investor later asks where the number came from, you click the card: the exact code run, its inputs and its delivery — provenance an auditor accepts.
The new ops hire
Week one every risky action needs a human click. By week three the colleague has earned autonomy on the three routine tasks it does cleanly — and only those. Compliance can rewind to each graduation.
The 2am incident review
Instead of grepping logs, drag the rewind bar to 2:14am: chat, wall and mood re-derive exactly as they were. You watch it think its way into the mistake.
The skeptical enterprise buyer
"How do we know your AI didn't fabricate its permission?" Every consent is an immutable event, a corrupted one freezes autonomy shut, and the chain replays in front of them.
06What is next
| next | in plain words |
|---|---|
| Vitto Live beside the chat | Watching the colleague think should happen where you talk to it, not on a separate page. Plus the Space wall naming which conversation it belongs to. |
| Snapshot bookmarks | Rewind stays instant no matter how many years of history accumulate. |
| The daily reflection | The colleague reviews its own day and writes what it learned into the diary — the step from reactive tool to colleague that improves. |
| The privacy switch | One control that stops memory, briefs and profiles together, for spaces that must be provably unremembered. |
Then real-time collaboration (Stage 2), which is where several people and the colleague share a space live. The foundations — members, groups, roles, per-workspace access — are already built and manageable in the interface today.
07Where to look next
| you want | go to |
|---|---|
| Live status, the plan, every founder decision | The Program Board |
| The engineering truth | docs/Project-analysis/EXECUTION-BOARD.md |
| Every known gap and its owner | docs/Project-analysis/GAP-LEDGER.md |
| To start working on it | docs/Project-analysis/2026-07-27-HANDOFF-next-session.md |
This briefing is generated from docs/Project-analysis/team-briefing.html in the agente repository and published to our own Cloudflare infrastructure. Edit the source, run npm run pages:publish, and this page updates. It is a companion to the Program Board, not a second source of truth — the repository is the truth, and both pages mirror it.