The workspace, mapped

The whole architecture on one page

What this is

This is the map of my AI workspace, the system I run my working day through. Every box on it exists and runs today. You talk to it from anywhere. One session does the work with everything loaded. A governance floor sits under every action. What a session learns persists on the left, what it produces ships on the right, and a watcher layer checks the system itself.

Map of the AI workspace: entry points, the session, the governance floor, persistent state, gated outputs, and the watcher layer. you keyboard · voice, desk or phone Interactive desktop app · terminal Remote phone · tablet · voice Scheduled six jobs on the OS clock 1 THE SESSION one agent, fully loaded loads at start rules & instructions typed memory lessons invokes skills: orient · board · wrap … roles ×17 → subagents workflows reaches out through MCP bridges: mail · calendar · browser git & GitHub 2 GOVERNANCE FLOOR hooks on every command · no credentials in files · outbound needs a human yes human gate 3 WHAT PERSISTS strategy task board lessons typed memory project context evidence ledgers 4 SHIPS deliverables & briefs public repos ×6 replies & posts next session starts already loaded 5 WATCHERS the system, checked weekly audit reads everything, files findings dead-man's switch flags any lane gone quiet security digest tripwires & floor blocks token meter spend & context telemetry backups + verify restic, off-machine → findings land on the task board, as work
  1. The session. Every piece of work happens inside one of these. It wakes up already knowing the rules, the current state, and every past correction, then does the task through skills, seventeen expert roles it can hand work to, and parallel agent workflows when the job is wide.
  2. The governance floor. Not advice, mechanism. Hooks screen every command before it runs, credentials live in a password manager the agent never reads, and nothing outbound moves without an approval, one item at a time.
  3. What persists. The part most setups skip. Strategy, the task board, lessons, typed memory, per-project context. Because this layer exists, no session ever starts from zero, and a correction made once holds forever.
  4. Ships. Client deliverables, six public repos, replies and posts. All of it exits through the human gate.
  5. Watchers. The system checks itself. A weekly audit reads the whole workspace. A dead-man's switch flags any scheduled lane that goes quiet, tripwires log what the floor blocked and a meter tracks spend, with verified off-machine backups under everything. What they find becomes work on the board.

Fourteen modules, seventeen roles, six scheduled jobs and a six-repo public mirror collapse to five boxes: one engine, a floor it stands on, state that persists, an exit with a human on it, and watchers over the whole. The complexity is quantity. The shape is simple.

Check my work

The architecture on this page is published. On GitHub, agent-workspace-architecture documents the patterns behind every box, extracted from the live system. If you want to know whether this is real, read it the way an auditor would.

18patterns published from this workspace
6public repos of its tooling
1human gate on everything outbound

One page, on purpose. The full version is the system itself.