> Paper: SkillOS: Learning Skill Curation for Self-Evolving Agents > arXiv: 2605.06614 > Team: UIUC · Google Cloud AI · MIT > Core claim: The bottleneck of agent skill libraries is not "adding skills" but "managing skills" — which to keep, which to delete, and which to iterate on.
Introduction: Your Agent Is Hoarding
A common approach for LLM agents playing multi-round games or doing complex reasoning is to give them a skill library: retrieve relevant skills for each task and execute them step by step. The fundamental problem: skill libraries only grow, never shrink, and become bloated.
Existing solutions rely on manual maintenance (exhausting), hardcoded rules (inflexible), or learn only insertion, never deletion (hoarder mode). The result:
- New skills get interfered with by old ones
- Redundant skills slow retrieval and dilute attention
- Low-quality skills reduce success rates
- Agent Executor (π_L): the frozen main model (Qwen3-8B, Qwen3-32B, or Gemini-2.5-Pro) that executes tasks. It never gets trained.
- Skill Curator (π_S): a small trainable model (default Qwen3-8B) that observes executor trajectories and decides
insert_skill,update_skill, ordelete_skill. - SkillRepo: an external Markdown-based skill store. Each skill has a name, description, workflow, disable conditions, and preconditions.
SkillOS's insight: skill management itself should be a learning problem, not a rule problem.
01 | Architecture: Decoupled Dual Modules
02 | Four-Part Composite Reward
Since curation effects are delayed, SkillOS designs a composite reward:
| Reward | Meaning | Weight | |--------|---------|--------| | r_task | success rate on subsequent tasks in the group | 1.0 (main signal) | | r_fc | format correctness of curation operations | 1.0 | | r_cnt | skill content quality judged by Qwen3-32B | 0.1 | | r_comp | library conciseness (1 - repo tokens/input tokens) | 0.05 |
Training uses GRPO with 8 rollouts per group and a dropped KL term to encourage exploration. The first task in each group runs with an empty library so rewards reflect curation decisions, not executor ability.
03 | Grouped Training for Delayed Feedback
Tasks are clustered by traits into groups: the first task's trajectory drives curation; later tasks validate it, with rewards backpropagated to the curation decision. Ablation: removing grouped training drops performance from 61.2% to 57.3% — the most critical component.
04 | Results: An 8B Curator Beats Gemini-2.5-Pro Curation
ALFWorld
| Executor | Method | Success | Steps | |----------|--------|---------|-------| | Qwen3-8B | No Memory | 47.9% | 21.1 | | Qwen3-8B | ReasoningBank | 55.7% | 20.1 | | Qwen3-8B | SkillOS | 61.2% | 18.9 | | Qwen3-32B | No Memory | 54.5% | 20.3 | | Qwen3-32B | ReasoningBank | 61.4% | 18.7 | | Qwen3-32B | SkillOS | 68.6% | 17.3 | | Gemini-2.5-Pro | No Memory | 66.4% | 17.7 | | Gemini-2.5-Pro | MemP | 74.3% | 15.2 | | Gemini-2.5-Pro | SkillOS | 80.2% | 14.8 |
Counterintuitive finding
Zero-shot curation by Gemini-2.5-Pro scores 79.3%, while a trained Qwen3-8B curator achieves 80.2%. Curation is a specialized, trainable skill — stronger reasoning does not equal better curation.WebShop + Reasoning
| Executor | Baseline | SkillOS | Gain | |----------|----------|---------|------| | Qwen3-8B | WebShop 33.3 / AIME 69.6% | 40.6 / 73.8% | +7.3 / +4.2 | | Qwen3-32B | WebShop 41.5 / AIME 74.0% | 49.2 / 79.7% | +7.7 / +5.7 | | Gemini-2.5-Pro | WebShop 48.6 / AIME 81.8% | 56.0 / 88.6% | +7.4 / +6.8 |
Curators trained on reasoning tasks (AIME/GPQA) transfer surprisingly well to agent tasks (ALFWorld), because reasoning-trained curation strategies are more abstract (decomposition, verification, adaptive planning).
05 | The Curator Evolves: From Hoarding to Refining
Operation distribution over training:
| Stage | Insert | Update | Delete | |-------|--------|--------|--------| | Early | ~80% | ~15% | ~5% | | Mid | ~50% | ~40% | ~10% | | Late | ~35% | ~50% | ~15% |
Library content evolves from generic tips to failure-handling logic, conditional branches, and systematic search strategies — emergent meta-skills. Skill usage becomes leaner and more precise: 1.95 skills per case vs 2.24 baseline; successful skill usage rises from 61.2% to 88.6%; skill coverage from 53.6% to 72.9%. The curator learns quality over quantity.
06 | Limitations and Industry Implications
Limitations 1. Training cost: 3-5 days on H100 GPUs 2. Task grouping relies on manual clustering; automatic grouping remains open 3. Executor is frozen; joint optimization is unexplored 4. Markdown format may limit expression of complex skills
Industry implications: Most agent frameworks assume static, human-maintained skills. SkillOS shows skill management can be automated, learnable, and evolving — personal assistants that add/remove skills based on your habits, enterprise agents that distill best practices, and game NPCs that learn new skills and forget bad habits.
Conclusion
SkillOS's core contribution is not "teaching agents new skills" but "teaching agents to manage skills." When a complex agent may hold hundreds of skills, deciding what to keep, delete, and iterate matters more than adding skill #101. Training a small RL-based curator outperforms both manual maintenance and zero-shot curation by the strongest models. If your agent's skill library is piling up in chaos, SkillOS offers an alternative: don't hire a maintainer — train one.
> References > - arXiv: 2605.06614 — Ouyang et al., "SkillOS: Learning Skill Curation for Self-Evolving Agents", May 2026 > - Code and data: not publicly released (as of 2026-06-30)