When code gets cheap, the bottleneck moves to evaluation. Not prompt engineering — everyone says that. Actual evaluation. Knowing which of the fifty functions your agent wrote is correct, and having the harness to verify it. That's where the hard work is in 2026.