What's this about?
A helper agent observes a specialised harness and corrects student actions in the simpler target harness, creating demonstrations for fine-tuning.
What exactly is harness distillation?
It teaches the model behaviours that previously depended on external orchestration, so the specialised runtime can be removed. Internalisation may reduce flexibility when workflows change.
Why does this matter?
Author-reported preprint; 10 HF upvotes at fetch time.
- Deployment can use one fixed harness across domains.
- Reported success exceeds the base model even with the old harness attached.
- Recovered patterns still need testing on unseen tools and models.