SkillOpt: Executive Strategy for Self-Evolving Agent Skills
Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Yuqing Yang, Dongdong Chen, Xue Yang, Chong Luo
arXiv:2605.23904 · 250▲ · cs.AI, cs.CL
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
SkillOpt is a text-space optimizer that treats an agent's natural-language skill document as a trainable object, borrowing deep-learning controls like batches, a textual learning rate (bounded number of edits per step), a held-out validation gate, and epoch-wise slow/meta updates. A frontier optimizer model reads execution trajectories, proposes structured add/delete/replace edits to the skill, and only accepts edits that improve validation performance, exporting a compact best_skill.md (300 to 2,000 tokens) while the target model and harness stay frozen.
Why it matters
For closed models where you cannot touch weights, this offers a way to adapt agent behavior by optimizing a plain-text skill file offline, then deploying it across models and harnesses without fine-tuning. The artifact is small, human-inspectable, and reusable, so you can audit what changed and transfer one trained skill to related models or execution loops.
Practical takeaway
If you run agents on tool-use or procedural tasks (spreadsheets, document QA, multi-step tool calls), you can iteratively train a reusable skill document instead of hand-writing prompts, and expect the biggest gains on procedural benchmarks and on smaller/weaker models. Watch for the validation gate: gains depend on having a held-out selection split to reject harmful edits rather than trusting the optimizer's self-critique.
Key result
Across 52 (model, benchmark, harness) cells, SkillOpt was best or tied-best on all 52. The headline case (GPT-5.5, direct chat) rose from 58.8 to 82.3 average over six benchmarks (+23.5 over no skill), and +5.4 over an oracle that picks the best of six competing baselines per cell; note the largest single gains (SpreadsheetBench 41.8 to 80.7) come from procedural tasks where the zero-shot baseline was weak, and all experiments use the same optimizer/target model family so results may not generalize to arbitrary setups.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.