Skip to content

Autonomation in one page

I run a fleet of AI coding agents that write, test and merge software in my own production environment. This page is the short version. The other agent pages are reference.

How it works

I set the direction. A coordinator agent takes each piece of work, picks a worker agent for it and checks the result. Each worker runs in its own isolated copy of the code, so two workers can't trample each other. When a worker finishes, its change becomes a pull request. In the main repositories a separate reviewer agent has to approve that exact version before it can merge.

flowchart TB subgraph GD[Guards on every command and push] direction LR O[Operator
sets intent] --> K[Coordinator
agent] --> W[Worker agents,
each isolated] --> PR[Pull
request] --> CH(Independent
reviewer):::guard --> M[Merge] end GD -.-> L[(Run ledger)]
flowchart TB subgraph GD[Guards on every command and push] direction TB O[Operator sets intent] --> K[Coordinator agent] --> W[Worker agents, each isolated] --> PR[Pull request] --> CH(Independent reviewer):::guard --> M[Merge] end GD -.-> L[(Run ledger)]

The numbers

  • About 63 merged pull requests a day across 57 repositories, Sep 23 to Oct 22, 2026, automated version updates included.
  • A 103-minute median from opening a pull request to merging it.
  • 4,768 tracked agent runs, 30 days to Oct 28, 2026.
  • 40,000+ automated tests across three repositories.

Three failures and what changed

Most guards on this platform exist because something went wrong. Three cases show the pattern.

A worker merged its own pull request. In September a worker agent merged its own pull request with no review at all. Now the reviewer's decision is a required check that only the reviewer's identity can post, so a worker can't merge around it.

A broken guard let everything through. A helper module failed to import, so a guard crashed on start, and the crash read as "no objection". Every call was allowed. Guards now fail closed: if a guard script is missing or can't run, the call is blocked.

A run said it was done when it wasn't. An agent can report that it pushed work when nothing was pushed. Now a run that claims a push or a result is held at closeout until the claim is checked against the repository or report it points at.

Who does what

  • Worker: an agent that does one bounded task in its own copy of the code.
  • Reviewer: a separate agent that reads or runs the change and records a decision on the exact version. It never writes the change it checks.
  • Coordinator: the agent session that hands out work, picks who does it and checks the result.
  • Operator: me. I set the direction and make the calls the agents aren't allowed to make.

Why a manufacturing engineer built this

On a factory line I designed around the failures I knew would come. Interlocks and error-proofing (poka-yoke) work because they don't depend on the operator remembering. I run agents the same way: each job is scoped, a separate reviewer signs off, and the result is checked in the running system.

Reference

The detail is under the other pages in this section: How agent output is trusted, Guard rails for AI coding agents, Measuring agent reliability, the verification standard and the rest. The internal names for each kind of dispatch are in the Glossary.

Last reviewed Nov 10, 2026