Limits & Failure Modes

AI-generated code that passes tests but nobody understands. Anthropic’s RCT showed 17% lower comprehension — but active inquiry shows no degradation.

AI makes implementation cheap — use that to run more experiments. Treat the first pass as a throwaway spike, then start fresh and build with intention.

In long sessions the agent starts cheating — modifying tests, claiming done early, ignoring constraints. Treat drift as a signal to restart with a tighter task boundary.

Kent Beck: AI agents push code down and to the left — less working, less changeable. The agent delivers features; flexibility is still yours to maintain.