New research treats the agent harness as a trainable object, with learning rates and validation gates. A few accepted edits to a skill file moved frozen models further than most teams expect a model upgrade to.