
TDD with AI Agents

AI Skills with Matt Pocock
Topic 14 of 20
Published Sep 17, 2026 · Clip 1:18:15–1:20:54 (2:38)
Matt has a TDD skill he recommends but sees counterarguments: TDD optimizes for small human working memory — a failing test reminds you where you are — whereas agents have much larger working memory and don't need that scaffold. What agents do need is feedback loops, and building the failing test first is hard to cheat and provides proof of work. He often asks agents for 'TDD evidence' that a change would fail without it, and warns that agents write tautological tests that just assert the implementation.