Stop Letting AI Mark Its Own Homework

Short summary

The author argues that AI coding tools like Claude Code produce plausible-but-subtly-wrong code because the same agent writes and verifies its own output, creating a circular trust problem. After evaluating GSD (a Claude Code planning tool), they built Greenlight, an open-source multi-agent system where separate agents handle architecture, testing, implementation, and security review. A mechanical test runner—not AI—serves as the final arbiter, ensuring code is actually correct rather than just appearing correct.

• •AI coding tools produce plausible code that passes glances but fails in production because the same agent writes and reviews its own work
• •Greenlight separates concerns across Architect, Tester, Developer, and SecOps agents with hard boundaries
• •A mechanical test runner (pytest, jest, go test) acts as the trust boundary—not any AI model

Generated with AI, which can make mistakes.