Red/green TDD - Agentic Engineering Patterns - Simon Willison's Weblog

Coding agents have two major failure modes: they write code that doesn't work, or they build unnecessary features that never get used. Test-first development solves both problems at once. The pattern is simple - write your automated tests first, confirm they fail (the red phase), then iterate on implementation until they pass (the green phase). This approach forces you to define what success looks like before the agent starts coding.

The red phase is critical and often skipped. If you don't confirm your test fails first, you might have written a test that already passes, which means it's not actually testing your new implementation. It's dead weight in your test suite. With agents, this verification step becomes even more important because the agent might misunderstand your requirements and write a test that checks the wrong thing entirely.

As projects grow, the risk of breaking existing features grows with them. A comprehensive test suite built through TDD is the most effective way to catch regressions before they ship. Every good model already understands "red/green TDD" as shorthand for this entire workflow. You don't need to explain test-driven development in detail - just tell the agent to use red/green TDD and it knows to write tests first, watch them fail, then implement the fix.