One control plane for every agent on my machine

When several agents are running, the work is not the hard part. Keeping track of them is.

One of them stops and asks a question. The question lands in a chat tab I am not looking at. An hour later the agent is still waiting, and the hour is gone.

Another one opens a folder it had no business opening. Nothing was lost that time.

I built one system so that stops happening.

One place for the work

Every agent on my MacBook runs under it. Claude, Codex, Kimi, OpenCode, all of them.

One database holds every task, every escalation and every session summary. When a session ends, its work does not end with it. The next agent reads the state from the database instead of from a conversation nobody can find again.

Every agent has a mandate

A mandate says what an agent may touch, and the operating system enforces those file permissions, scoped to what that one task could damage.

An agent that needs more access asks its manager, or asks me. It does not take it.

Questions reach me, not a chat tab

When an agent needs a decision, the escalation reaches me at the operating system level. A card appears on my screen with the question and the real choices. I answer there, and the agent goes on.

That is the part that changed my days most. A decision that used to sit unread now takes a minute.

What it has carried

On my own machine this summer: several hundred manager sessions and about nine thousand worker sessions.

I did not plan for that number. It arrived because starting an agent keeps getting cheaper, while losing track of one costs the same as it always did.

Why I keep working on it

This is my own working system. I use it every day, and I keep changing it because I keep finding the next thing that breaks at this scale.

This is the problem I work on.