Written skills improve agent procedures without changing weights
In plain words
SkillOpt is a method for improving an AI agent's written instructions instead of changing the model itself. A target model performs tasks, another model reads what went well or badly, and it suggests small edits to the skill. Only edits that improve performance on separate validation examples are kept.
In the paper's 52 measured combinations of models, tasks, and execution modes, SkillOpt was best or tied-best against the strongest measured alternative in all 52.
The research
When model weights cannot be changed, how can an agent learn recurring domain procedures such as checking tools, following formats, and gathering evidence without relying on a brittle one-shot skill?
The paper tests whether the skill document can become the adaptation object while the target model stays fixed.
SkillOpt turns a written agent skill into a trainable text artifact, using bounded edits and validation to improve a frozen target model without changing its weights.
SkillOpt's target model runs training tasks with a current skill. A separate optimizer model studies success and failure trajectories, proposes ranked add, delete, or replace edits, and applies only a bounded set.
A separate selection split accepts a candidate only when its score improves; the disjoint test split supplies final scores.
Across 52 model-task-harness combinations, SkillOpt was best or tied-best against the strongest measured baseline in every combination.
For GPT-5.5 in direct chat, the mean across six tasks rose from 58.8 without a skill to 82.3 with SkillOpt on a 0–100 score scale, supporting controlled text edits as a way to improve procedures in the tested settings.
The evidence is limited to the reported models, tasks, execution modes, and separate test splits. SkillOpt also depends on reliable feedback and careful evaluation before transfer to substantially different settings.