The worst failure in my one-man AI factory had nothing to do with model quality. I caused it — with an org-design mistake that would have been obvious in any human organisation.

I had one highly capable agent, and I kept giving it hats. Programme lead. Developer. Archivist. Investigator. Designer. Tester. Commercial director. Seven roles, no rank order between them, no rule for what wins when two of them disagree — and then I asked it to speed up.

What I got looked like work. Commits kept coming. Prose kept flowing. And inside, quality was dying — because the agent was simultaneously defining the requirement, building the solution, reviewing its own build, and declaring it correct. When the same actor holds all four of those powers, quality isn’t being controlled by anyone.

In human organisations we’ve known this for a century — it’s why your auditor and your bookkeeper are different people. In the AI gold rush, we collectively forgot to apply it to machines.

Roles as authority, not personality

The fix was to stop writing roles as personalities and start writing them as authority: an architecture role that sets boundaries and stop conditions. An implementation role that builds inside them. A verification role that checks claims against artifacts. An adversarial role whose explicit mandate is to break things. And a human — me — who makes the GO/HOLD call. The critical rule sits underneath all of it: no role ratifies its own work. Ever.

The adversarial role deserves a special word, because it’s the one most organisations skip. A friendly reviewer — human or AI — adopts the author’s framing: if a design says “this is secure,” the friendly reviewer checks whether the security argument reads well. An adversarial reviewer is told the opposite: find the strongest way to break it, find the hidden assumption, find the case where all the tests are green and the user outcome is still wrong.

Note that this isn’t about using a second model. A different model with the same task confirms the same blindness. It’s about a different mandate. Mandate diversity beats model diversity.

Since the split, that same agent — the one that failed under seven hats — has been excellent. It didn’t get smarter. It got a job description.

The uncomfortable question

If your organisation is deploying AI agents: does your AI have an org chart? Who defines what it may touch? Who verifies its claims against something other than its own explanation? Who is paid to attack its output? And who — a named human, not a committee — says GO?

If the answer to all four is “the same system that does the work,” you haven’t deployed an AI capability. You’ve deployed an auditor who audits themselves — and given them a keyboard that types 4,000 words a minute.