• Matt has a TDD skill and still recommends it, but he has become more mixed on how it fits agents. 1:18:15
  • TDD helps humans by keeping working memory small: one failing test tells the developer where to resume. 1:18:35
  • Agents have larger working memory than humans, so TDD does not solve exactly the same problem for them. 1:18:47
  • Agents still need feedback loops, probes, and evidence that their changes work in the code environment. 1:19:08
  • Having an agent build the failure first makes it harder for the agent to cheat and gives the human proof that the change mattered. 1:19:36
  • Matt often asks for "TDD evidence," meaning proof that the behavior would fail without the change. 1:19:36
  • A common failure is tautological tests, where the test merely repeats the implementation. 1:20:06
  • Matt defines tech debt as anything that makes the codebase harder to change over time. 1:21:30
  • A good codebase is easy to change without causing cascading failures. 1:21:30
  • Automated review can help by checking coding standards and catching poor tests, but the review agent's quality also needs evaluation. 1:21:49
  • Agents did not remove the old tech debt problem; they made the same problem faster and more visible. 1:22:48