A quest to hand me?
If you have put agents to work and you no longer know who triggered what, whether it was allowed, or how to prove it — that is the problem I work on every day.
What I can do
Put a gate in place that refuses in code. Not a rule written in a document nobody opens: a gate the machine enforces, that returns the same verdict for the same request, and writes down why it said no. On my own system, 2,156 crossings are recorded — and almost one in four is a refusal.
Have a machine review without giving it the right to write. An observer that returns a reasoned opinion and can change nothing. The lock is not an instruction, it is its list of tools.
Measure who really knows a codebase. I re-read 25,039 commits across two open projects to show that the usual measure gets it wrong: it counts email addresses instead of people, and sometimes names an owner who does not exist. The method is published, with its independent oracle.
Make a failure loud. A stopped backup, a refused push, a silent agent: as long as nothing says so, silence looks like everything working.
Proof, not promises
- The full study — the method, the figures, and what they do not prove.
- Fourteen open projects — readable, runnable, reusable.
- Seventy-seven articles — many of them about an outage and what it cost.
Who for
A team putting agents into production that wants to sleep at night. A company that must prove to a client or an auditor that its agents stayed in scope. An open project looking for someone to put its guardrails in place.
How it works
A free half-hour first: you describe what worries you, I tell you whether I can handle it — and if I cannot, I say so too.
Get in touch
Write one sentence about what is blocking you. I answer everyone.