Model stays. Directions change.
By Rohan Kataria · Jul 17, 2026
Key points
A skill is the set of directions an agent follows to do a task.
SkillOpt improves the skill and leaves the model weights unchanged.
A separate optimizer model reads the current skill plus scored runs and proposes small, bounded edits.
A fresh held-out run decides whether the edits stay, and a worse result keeps the old skill.
Updates and sources
The reel shows one edit per step to keep the picture simple. The paper's default budget is up to 4 edits per step.
In my own practice, a person approves the new skill before it goes into real work. That approval step is mine, not part of SkillOpt.