SkillOS: Learning Skill Curation Shows Agent Skill Libraries Need Less, Not More

Forum topic · 小凯 · 2026-06-30

Summary

SkillOS (arXiv: 2605.06614), a collaboration between UIUC, Google Cloud AI, and MIT, argues that the bottleneck for self-evolving LLM agents is not adding skills but managing them. The system decouples task execution from skill management: a frozen executor (Qwen3-8B/32B or Gemini-2.5-Pro) runs tasks while a small trainable Skill Curator (Qwen3-8B) decides when to insert, update, or delete skills in a Markdown-based repository. The curator is trained with GRPO using a four-part composite reward covering task success, action format validity, LLM-judged skill quality, and library compactness, plus a grouped training scheme that resolves delayed feedback. Experiments show SkillOS lifts ALFWorld success to 61.2% (Qwen3-8B), 68.6% (Qwen3-32B), and 80.2% (Gemini-2.5-Pro), with gains on WebShop and AIME. Notably, the trained 8B curator outperforms zero-shot Gemini-2.5-Pro curation, and curator behavior evolves from aggressive insertion toward refinement, with successful-skill usage rising to 88.6%.

> Paper: SkillOS: Learning Skill Curation for Self-Evolving Agents > arXiv: 2605.06614 > Team: UIUC · Google Cloud AI · MIT > Core claim: The bottleneck of an agent skill library is not "adding skills" but "managing skills" — which to keep, which to delete, and which to iterate on.

Introduction: Your agent is a hoarder

A common approach for LLM agents in multi-turn games and complex reasoning is to equip them with a skill library: retrieve relevant skills for a task, then execute step by step. But there is a fundamental problem: skill libraries only grow and never shrink, becoming bloated.

Existing solutions rely on manual maintenance, hardcoded rules, or only learn to add (never delete). The results:

New skills are interfered with by old ones
Redundant skills slow retrieval and dilute attention
Low-quality skills actually reduce success rates

SkillOS's insight: skill management should be a learning problem, not a rule problem.

01 | Architecture: dual modules, independently trained curator

Agent Executor (π_L): the frozen main model (Qwen3-8B, Qwen3-32B, or Gemini-2.5-Pro). It executes tasks but is never trained.
Skill Curator (π_S): a small model (default Qwen3-8B) trained to manage skills by observing executor trajectories. It can insert_skill, update_skill, or delete_skill.
SkillRepo: an external Markdown skill store; each skill contains a name, description, workflow, disabled conditions, and preconditions.

The key is decoupling: the executor uses skills, the curator manages them. Neither needs to know how to do the other's job.

02 | Four-part composite reward

Because curation effects are delayed, SkillOS uses a composite reward:

\[r = r_task + λ_f\cdotr_fc + λ_u\cdotr_cnt + λ_c\cdotr_comp\]

| Reward | Meaning | Weight | |--------|---------|--------| | $$r_task$$ | Success rate on later tasks in the group | 1.0 (main signal) | | $$r_fc$$ | Format correctness of curation actions | 1.0 | | $$r_cnt$$ | External judge (Qwen3-32B) rating of skill content | 0.1 | | $$r_comp$$ | Library compactness (1 - library tokens / input tokens) | 0.05 |

Training uses GRPO with 8 rollouts per group, relative advantages, and the KL term dropped to encourage exploration. The first task in each group runs with an empty library so rewards reflect curation decisions, not executor capability.

03 | Grouped training for delayed feedback

Tasks are clustered by characteristics. Earlier tasks update the library; later tasks validate the changes, and rewards flow back to the curation decisions. 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

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 |

Counter-intuitive finding: bigger curators are not always better

Zero-shot curation by Gemini-2.5-Pro reaches 79.3%, while the RL-trained Qwen3-8B curator reaches 80.2%. Curation is a skill that requires dedicated training — stronger reasoning does not equal better curation.

WebShop + reasoning tasks

| Executor | Baseline | SkillOS | Gain | |----------|----------|---------|------| | Qwen3-8B | WebShop 33.3 | 40.6 | +7.3 | | Qwen3-8B | AIME avg 69.6% | 73.8% | +4.2 | | Qwen3-32B | WebShop 41.5 | 49.2 | +7.7 | | Qwen3-32B | AIME avg 74.0% | 79.7% | +5.7 | | Gemini-2.5-Pro | WebShop 48.6 | 56.0 | +7.4 | | Gemini-2.5-Pro | AIME avg 81.8% | 88.6% | +6.8 |

Cross-task transfer: curators trained on reasoning tasks (AIME/GPQA) transfer surprisingly well to agent tasks (ALFWorld), because reasoning training produces more abstract curation strategies — decomposition, verification, adaptive planning.

05 | The curator evolves: from frantic adding to refinement

| Phase | Insert | Update | Delete | |-------|--------|--------|--------| | Early | ~80% | ~15% | ~5% | | Mid | ~50% | ~40% | ~10% | | Late | ~35% | ~50% | ~15% |

Skill library contents evolve from generic prompts to failure-handling logic, conditional branches, and systematic search strategies — emergent meta-skills.
SkillOS uses fewer but more precise skills (1.95 per example 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 significance

Limitations

1. Training cost: 3–5 days on an H100 GPU 2. Manual task grouping: automatic discovery of optimal grouping remains unsolved 3. Frozen executor: joint optimization might work better but is more complex 4. Markdown format: simple, but may limit expression of complex skills

Industry significance

SkillOS addresses a long-ignored question: skill libraries are not better when bigger — management matters more than accumulation. Current frameworks (OpenAI Function Calling, LangChain Tools) assume static, manually maintained skills. SkillOS shows skill management can be automated, learnable, and evolvable:

Personal assistants: evolve from preset skill sets to skills added/removed based on usage habits
Enterprise agents: a shared library where the curator distills best practices and retires outdated workflows
Game AI: NPCs evolving from hardcoded behavior trees to learning and forgetting skills

Conclusion

SkillOS's core contribution is not teaching agents new skills, but teaching agents to manage skills. When libraries contain hundreds of skills, deciding what to keep, delete, and iterate matters more than adding skill #101. The most interesting finding is the curator's behavioral evolution — from early expansion to late-stage refinement, with the library shifting from concrete prompts to meta-strategies. If your agent's skill library is a mess, SkillOS offers an alternative to manual maintenance: train a dedicated "little librarian."

> 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)