The biggest problem I’m seeing in AI right now is sycophancy: models agree with you (and with themselves) way too easily. And it hits hardest in AI coding, because the worst thing you can do is have a coding agent evaluate its own work after it’s been swimming in your opinions for an hour. It’s like a student grading their own homework—maybe it flags a couple cosmetic issues, but the real bugs and bad decisions get stuffed under the rug until your codebase is on fire. In this video I show the fix: give your builder a sparring partner. I break down Anthropic’s new GAN-inspired harness for long-running coding sessions—planner + generator + evaluator—where the evaluator’s sole job is to be nitpicky and rip apart the implementation. Then I get practical and walk through my own “adversarial dev” harness that I built from scratch (Claude Agent SDK + Codex SDK, and you can even mix them). The key is sprint negotiation before any code gets written: the generator and evaluator agree on phases, scoring criteria, and thresholds, then iterate (up to retries) until each sprint passes. Yes, it costs tokens, but the reliability boost is worth it—and it’s how you can use smaller/faster models and still ship surprisingly polished apps.