TDD still works for AI agents, with a few changes. A walkthrough with an open-source multi-agent example: requirements, failure modes, experiments and CI.