Autonomation¶
I make the agents reliable the way I would make a machine safe.
Status: in daily use · checked Oct 29, 2026
In plain terms
One system runs my AI coding agents. It stops itself when something is abnormal. Recorded failures drive new guards and regression checks, with enforcement limits documented.
Start with two pages: How agent output is trusted and Guard rails for AI coding agents. If you hire for AI infrastructure, read Measuring agent reliability next. Everything else is reference.
Where to start¶
Pick the track that fits you; each page reads on its own.
- Hiring for AI infrastructure or developer productivity: how agent output is trusted, guard rails, measuring agent reliability, then what a run leaves behind.
- Hiring for delivery or forward-deployed work: how I run AI coding agents, then what a run leaves behind.
- Hiring for robotics or manufacturing engineering: start with Hardware and manufacturing, then the verification standard, which brings commissioning discipline to software changes.
AutonomationThe name I use for the whole system: the agent fleet, its guards, the control plane and the review gates. It is the English term for Toyota's jidoka: automation that stops itself when something is abnormal. Glossary is the name I use for the whole system: the agent fleet, its guards, the control plane, and the review gates. It's the English term for Toyota's jidoka: automation that runs on its own, stops itself when something is abnormal, gets its root cause fixed by a person, and has that fix built in as a guard or a regression check.
I run a fleet of AI coding agents. I built the platform and set its standards and acceptance criteria. I decide which agent gets which work, what it may touch, who checks it, and what changes after a failure. In the main repositories, every pull request must pass an independent AI review gate pinned to the exact commit, with an extra adversarial review for security-class changes. Separately, recorded failures drive new guards and regression checks, with enforcement limits documented.
The system in one picture¶
sets intent] --> K[Coordinator
agent] --> W[Worker agents,
each isolated] --> PR[Pull
request] --> CH(Independent
reviewer):::guard --> M[Merge] end GD -.-> L[(Run ledger)]
| Part | What it does | Read |
|---|---|---|
| The fleet | Claude Code, Cursor, Codex and open-source models do the work. A coordinator agent hands it out through five kinds of lane, and every worker gets its own isolated checkout. | Lanes and the run ledger |
| Guards | Hooks and git-level backstops intercept risky commands and pushes. Each one exists because of a real failure, and each is checked to still be doing something. | Guard rails |
| The control plane | A ledger records every dispatched run. Sessions address each other by role. Multi-step plans have owned paths and gates between phases. | Lanes and the run ledger |
| Review gates | Every pull request in the main repositories must pass an independent AI review gate, and closing a task needs evidence. | How agent output is trusted |
What it has done¶
Each figure below carries its window and its source; how the figures are checked is on Site notes. The rows that matter most are the merged pull requests a day and the agent runs; the rest describe the size of the environment.
| Claim | What it means | Window and source |
|---|---|---|
| about 63 merged pull requests a day | A daily average over the window, so quiet days and busy days are blended; automated version-update pull requests are included. | Sep 23 to Oct 22, 2026. merged pull requests across my private repositories, automated version-bump pull requests included, averaged per day. |
| 4,768 agent runs | Every dispatched session is a row, including read-only review jobs and failed or cancelled runs. Interactive sessions and scout investigations aren't rows. | 30 days to Oct 28, 2026. dispatched agent sessions recorded in my run ledger. |
| 40,000+ automated tests across three repositories, 20,000+ of them in the main repository | A count of tests, not a measure of how much they cover. Collected, not passing. | As of Oct 29, 2026. tests the test runner collects in the three repositories that hold my automation code, counted together. |
| 73 architecture decision records | Written decisions, each with the options I rejected and the trade-offs I accepted. | As of Oct 2026. decision records written for the environment. |
| 9 main machines | A count of hosts, not a measure of capacity. | As of Oct 2026. main hosts in the environment I run. |
| 96 private repositories | A count of repositories, not of services. | As of Oct 29, 2026. repositories in my private GitHub organization. |
How often agent runs finish, fail or stall is on measuring agent reliability. Where each part of the system stands is on How agent output is trusted.
The rest of the set¶
| If you want to see | Read |
|---|---|
| how multi-step changes are split across workers and checked | Orchestrating multi-step work |
| how often runs fail, and what catches each kind | Measuring agent reliability |
| how model cost is routed and capped | Routing agent work by cost |
| the layers underneath, and how a new service is added | The platform path |
| the artifacts a run leaves behind, in redacted form | What a run leaves behind |
| how the environment is observed, and how alerts are designed | Observability and alerting |
| trade-offs and the options I rejected | Selected design decisions |
| how documentation is kept true | Documentation that stays true |
The ideas that recur¶
- Guards, not guidelines. If an instruction matters, a hook, a gate or a test enforces it. A document is the weakest level, and pages say which level backs each rule.
- Read-only by default, reversible by design. Looking is cheap and needs no write access. Changes go through isolated checkouts and review gates, and a self-canceling rollback comes before anything hard to undo.
- Cost is an engineering constraint. Work goes to the cheapest model tier that can do it, and the expensive tiers need a stated reason.
- A check has to be able to fail. A guard, an alert or a test that has never been seen to refuse anything is a hypothesis.
This is a public edition: it leaves out anything private to the environment or the people around it. See About this site.