Disciplining the Coding Agent

Four posts on the same problem from four angles. If you did not write the code and cannot read all of it, what is left to trust?

The answer each post converges on is: not the code, and not the model’s account of itself, but the environment — a contract the agent cannot renegotiate, a decision rule that was allowed to say no, an answer key sealed before the first line was written, and a standing assumption that every one of those guards is dead until something has tried to kill it.

The origin of the whole discipline is a project that had none of it: The Ablation That Beat the Agent, where holding both pens — designing the benchmark and the thing being benchmarked — produced a headline number that did not survive a fair re-run.