The second-most common failure (after one-shot gambling) is running a beautiful loop with no definition of done.

The agent always reports success. That is its job. Yours is to disagree.

Symptoms

  • “Looks good to me” is the only test
  • No trace
  • Retry that rewrites the spec downward
  • Summaries instead of files
  • Nobody clicked the links

Fix

Write failing tests first in the spec. Even three rows in a table. Eval.

If you cannot think of a test, you cannot think of a job. Stay in chat.

The social version

Teams celebrate demos. Demos have no eval. Put eval in the same meeting as the demo or you will ship the demo.