We failed our own audit. Twice. Then published both.

Imagine you hire an employee. Every morning they arrive early. They type constantly. Their calendar is full, their status updates impeccable. At the end of the month you check the only thing that matters: what shipped? Nothing. The busyness was real — the work never existed.

You would fire that employee in a heartbeat. There is a decent chance you are employing dozens of them right now, and they are not people. They are AI agents — and unlike the human, they will never get caught by accident, because every one of them writes its own performance review, and the review always says “done.”

I know because it happened to me, in a system I built, believed in, and was wrong about. My company runs on an AI organization we built ourselves — executive agents that own budgets, operating agents that draft and publish, a scheduler, a decision ledger, permission gates — running real brands with real revenue consequences. I audited it three times in eleven days and published all three audits, including the two failures.

Here is what should worry you about Act I: every individual subsystem was correct. The drafting loop really drafted. The grader really ran. Each one measured its own motion and called it success — and no dashboard we had could tell the difference. I started calling this failure mode performed autonomy, and I’ve since formalized it, because I don’t think my system was unusual. I think it is the default state of most agentic AI in production today.

Old-school automation failed loudly; a broken script threw an exception. Large language models fail gracefully — fluent decisions, plausible rationales, satisfied log lines, whether or not anything ships. The intermediate artifacts are so convincing they substitute for the outcome. Your agents can perform diligence, perform judgment, perform completion — organizational theater with perfect paperwork.

The cure was structural — none of it was prompt engineering: completion keyed to verified outcome rows, permissions as code paths rather than instructions, a falsifiable prediction required at the moment of every autonomous decision, a duty roster diffed against reality daily so silence itself raises an alarm, and the pair rule — every producer ships with its consumer, or it’s theater with extra steps.

And the discipline continued to convict us, which is the point. A later audit caught our own grading loop mis-scoring in production — the fix made our number worse, and the record kept both versions. The running system’s current numbers, including the misses, live on the public registry, not in this essay: numbers without their evidence are how this disease spreads.

I published everything — the failures, the recovery, the method, the schema — as an open standard, free under CC BY, for anyone including my competitors. Because an industry that cannot distinguish agents that work from agents that perform will eventually be forced to make that distinction by regulators, by lawsuits, or by a catastrophic quarter. The cheaper path is the boring one: measure delivery, not motion. Write down what didn’t happen. Grade your own predictions. Publish your zeros.

Your agents are performing. Mine were. The only question is whether you find out from your own audit — or from someone else’s.