Confidence: medium. Evidence: benchmark; limited long-run production. Last substantive change: 2026-07.
This subsystem owns how the factory gets better over time: not the model's weights, but the prompts, context, tools, workflows, and harness code that shape how work gets done.
The conclusion
Improve the process that generates changes, but keep that improvement process inside a stricter outer control loop. Every mutation to the harness should be versioned, predicted, evaluated against frozen evaluations, rolled out to a limited slice, and automatically reverted if it regresses. The near-term target is a self-improving harness, not a self-improving model.
How the thinking got here
Humans tuned prompts by hand, then optimizers tuned prompts and skills, then agents began editing workflows and harness code, and then governed mutation with transferable harness evolution emerged. The consistent lesson is that self-improvement is plausible at the harness layer precisely because the harness can be versioned, evaluated, and rolled back in ways a model cannot.
Credible alternatives, and when each is right
| Approach | Right when |
|---|---|
| Manual harness engineering | small scope, high-stakes changes |
| Offline search | improvements can be explored safely first |
| Bandits | choosing among known variants online |
| Evolutionary search | large design space, cheap evaluation |
| Self-rewriting harness | strong gates and frozen evaluations exist |
Where it fails and what we still don't know
Failures include evaluator drift, overfitting to the evaluation, diversity collapse, and unsafe capability gain slipping past weak gates. Benchmark evidence is promising; long-run production evidence is limited. Open questions include rollback of learned state, promotion governance, and long-term repository health under continuous mutation.
What would change our mind
Sustained production evidence that a governed, self-improving harness raises accepted-outcome quality over many months without evaluator drift would make this a core capability rather than an experiment.