AI Office

Running several autonomous workers at once turns into tab-switching and guesswork. Logs tell you what happened; nothing tells you what's happening. Trust dies when state is invisible.

Anyone directing multiple agents on real tasks who needs to know, at a glance, who's thinking, who's working, who's blocked, who's done.

1. Space over lists.

   A 3D room instead of a log feed — position and posture encode agent state pre-attentively, so you read the room the way a manager reads an office, not the way an operator reads a terminal.

2. State is the interface.

   Four states only — thinking, working, blocked, done. Resisting a richer status taxonomy kept the room readable and the orchestrator honest.

3. Steerable, not just watchable.

   Every agent can be interrupted and redirected from inside the room; visibility without control is a dashboard, not an office.

A monorepo: packages/shared holds the typed event contracts, apps/server is the Node/TypeScript WebSocket orchestrator, and apps/web is the React + react-three-fiber office. The 3D layer renders purely from the event stream.

The engineering decision that paid off: the office was built first against a fake event emitter, then real Claude Agent SDK execution was swapped in behind the same contract — without touching the 3D layer.

Shipped so far: typed TypeScript event contracts shared by server and UI — the spine of the system; the 3D office rendering live employee states from a WebSocket stream; and real agent execution via the Claude Agent SDK — one subscription credential powers every employee, no per-model API keys, with built-in web search/fetch and external MCP tools. 214 tests passing.

The roadmap: roster, persistence, and parallel employees. Then a shared message bus, so a goal lets the team self-organize. Then the headline — record a good run and save it as a one-click replayable workflow. Then Docker sandboxing.