Why AI Agents Should Not Grade Their Own Work: Lessons on Feedback Loops and Parallelism

A writer who produces reading notes on technical books uses AI agents to proofread them, but found that reviewers flagged accurate facts as invented because they were given condensed notes instead of the source text. The author argues a judge can only assess what it can see, and that an agent scoring its own output tends to rationalize its mistakes. The piece describes feedback loops in which one agent writes, a separate agent scores, and corrections repeat until a threshold is met. It also covers running several agents in parallel, including subagents that each keep their own context and return only a final message. The author draws on two 2026 books, one on agentic architectural patterns and one on spec-driven development, which frame these ideas as design patterns and warn about failure modes such as collusion between generator and judge.