The home agent fleet
An always-on team of AI agents that runs on a Mac Studio in my apartment.
- Status
- Running since spring 2026
- Year
- 2026
- Role
- Solo
- Area
- Agents & AI systems
In plain English
I wanted AI agents that do useful work on a schedule without someone watching every step. The fleet runs at home on my own hardware and reports back to a private chat room. It is my testbed for making unattended agents dependable.
How it works
- Scheduled jobs start agents automatically using launchd (the Mac's built-in task scheduler).
- Each agent builds in its own git worktree (an isolated copy of the code), so experiments never collide.
- One agent writes the work and a different agent reviews it. Nothing merges on a single agent's say-so.
- Every loop has a kill switch and a spending cap.
- Results go to a Matrix room (a private chat server) on a homeserver I run myself.
What I did
I designed and run the whole setup myself, from the scheduling and review rules to the safety limits.
What I learned
Six months in, the reliability engineering has been the most valuable part. I wrote up what worked and what stayed ordinary in "Six months of agents in my apartment".