Suraj Sharma (@suraj_sharma14) posted a list of projects he says an AI agent engineer must build. This is the first of three replies: a teardown of my own setup. Here is how my agent setup is laid out, and what I cut from it on Oct 19.
Two sides
Building happens in High Command, my design system for building with agents. It runs as lanes of coding agents. One lane runs operations. Merging and deploying happen only on my explicit yes. One lane reviews anything where the truth matters. One holds the workers that make the changes. One keeps watch.
The other side is a set of chat agents, each with one job. On the evening of Oct 19, I named one of them, Firstmate, as the front door. It advises me and keeps the chat-side agents working in line with the build side, as far as it can. It is not a separate authority. Approval comes from me.
What I found
There were 42 agents. About 10 of them existed only to watch the same two outcomes. Some did nothing but summarize the other watchers' reports. That is seats reporting on seats.
The irony was the seat built to catch this. Its job was to flag when the fleet was mostly reporting on itself. Every run it made came back UNMEASURED. The inputs it needed had never existed.
And the one seat that would have tested an outcome directly was never built. It has been a proposal since September 6. Ten watchers got built instead.
How it happened
The watchers went up in one evening, as a time-boxed experiment. Each one had a "pause unless renewed" date. Nothing enforced the date. Several kept running for weeks.
What I cut
On Oct 19 I approved the cut: of the 10 watchers, 2 stay. They are:
A live page probe. It fetches the pages people actually get and flags failures. That is evidence from the wire.
One roll-up. It asks whether the two outcomes moved and names what is blocking them. It also writes the board that the old summary seats used to write.
Of the other eight, one had already been switched off in September. The other seven are retired now. Their records stay readable.
Then I fixed the registry, the list of which agents exist. Every agent now has exactly one row with a real status: 42 folders, 42 rows, no duplicates, no orphans.
What I don't have yet
A circuit breaker. A routine that keeps failing keeps running until someone notices. That is the gap I'd close next.
What I'd keep from this
- Count watchers against builders. If the watchers win, cut.
- A seat whose only inputs are other seats' reports is overhead.
- Give every experiment an end date, and something that enforces it.
- Cutting work that doesn't move the outcome is a result, not a retreat.
Limits
One builder's setup, counted on Oct 19, 2026. Whether two watchers are enough is not proven yet.
How many of your agents only watch other agents?
Cutting work that doesn't move the outcome is a result, not a retreat.