The AI stress-tested its own interface

The bottleneck in agent-written software was never generation — it was verification. Code that looks convincing and behaves badly is the default failure mode, and a human reviewing the diff catches less of it than they think.

Here the agents were pointed at their own output with instructions to find faults. That closes the loop: the same capacity that produced the bug is turned on finding it, which is a far better use of it than a reviewer's attention.

This is the single most practically useful pattern in the whole set. Everything else is a demo; this is a working method.