Mine production traces to train cheaper specialized AI agents

Original: Build agent -> get traces -> write evals -> hillclimb/rl -> better agent & harness

Source: x.com ↗