Harness Week, day three: agent-written code needs review more than human code, not less, and it arrives faster than humans read. Build a working LLM self-critic in ten lines, then swap in Macroscope’s AST-aware review for the production loop, and the human keeps the only verb that matters: merge.