The agent system I run my own work on.

I designed the system I work in every day and built it with AI agents. My notes, rules and project status live in one place and sync across my machines.

I built it so that agents read my email, Monday, Slack, calendar and meeting notes, and turn what needs doing into one task queue with a dashboard. While my main machine is on, a guard saves a snapshot of the notes every five minutes: more than 12,000 since August. If two sessions overwrite the same file, I can usually get the earlier version back. In one week in September, a review of the guard's alerts found four edits that sync had overwritten, and all four came back from the journal.

I'm also piloting a room where Claude Code, Codex and two other agents work on the same task. The idea is mine: the agents don't just answer side by side, they criticise each other's work. One case from August: a finished task wouldn't close, partly because Windows re-encoded a few accented letters, so two copies of the same output no longer matched. One agent reviewed the proposed fix, the agents implemented it, the tests passed and the task closed. The room runs as a pilot on one machine for now, and I keep extending it.

It's the same approach I bring to a client. The agents take the repetitive work, and the checks stay in.