From the very first commit of my Python project, one rule mattered more than any other: the coding agent must never see the acceptance criteria. So I built two agents that never compare notes: one writes the tests from the full specification, and a separate coding agent writes the code from the same specification with the criteria cut out. In most coding environments, the same agent writes both, which compromises the validity of the test results. I forced the cut to happen in code by using context assembly , the step that builds the coding agent’s prompt, where the criteria section of the task file is cut out before the prompt is built.

For a while I thought that cut was airtight. Then I found that a single comment line broke it: the cut stopped there, and every criterion after the comment reached the coding agent. A week later, a criterion wrapped onto a second line did the same thing. Not good. Both bugs were mine and I was annoyed that the same core problem had resurfaced.…