:→ A Coder agent writes the actual implementation
→ A Reviewer agent — a completely separate AI persona — reviews it for correctness, edge cases, and security issues
→ If rejected, the Coder revises based on specific feedback (up to 3 rounds)
→ Once approved, a Tester agent writes real pytest tests from scratch
→ Tests run in a sandbox. If they fail, the Coder fixes the code and the cycle repeats