SkillOS: Enabling Agents to Self‑Manage Their Skills

Paper Overview

The paper SkillOS: Learning Skill Curation for Self‑Evolving Agents (arXiv:2605.06614) proposes that skills act as an agent’s “procedural memory”. Historically, skills are hand‑written and manually maintained; SkillOS introduces a trainable “Skill Curator” that autonomously creates, updates, and removes skills.

System Architecture

The system consists of two parts:

Agent Executor (frozen) : Executes tasks by selecting relevant skills from a SkillRepo and running them.

Skill Curator (trainable) : After each task, observes the execution trace and decides whether to insert , update , or delete entries in the SkillRepo.

Problem Statement

Current LLM agents treat each task as a one‑off episode, forgetting past experience. Existing approaches—manual skill authoring, heuristic rules, or short‑horizon RL—fail to provide scalable, adaptable skill management.

Key Design Choices

Task‑flow grouping + two‑stage evaluation : Tasks are grouped by skill relevance into streams; early tasks update the SkillRepo, later related tasks evaluate the usefulness of those updates, supplying delayed reward signals to the Curator.

Composite rewards : Because downstream correctness alone cannot be attributed to a specific skill edit, the authors combine multiple reward components to more precisely credit each skill operation.

Markdown skill format : Skills are stored as Markdown files, aligning with Anthropic, OpenAI, and Hermes Agent skill formats, facilitating migration, human readability, and LLM generation.

Experimental Results

SkillOS consistently outperforms both memory‑free baselines and strong memory‑based baselines—achieving higher speed and accuracy.

The trained Curator transfers across different Executor backbones (i.e., it works when the underlying model changes).

The Curator also transfers across task domains.

During training, the SkillRepo spontaneously develops higher‑level “meta‑skills”, indicating emergent hierarchical structure.

Why the Paper Matters

1. It casts skill management as an RL problem, providing the first feasibility proof for learning skill‑curation policies.

2. The choice of Markdown unifies emerging industry standards for skill representation.

3. It offers a concrete mechanism for “self‑evolving agents” by anchoring evolution to the Curator’s ability to manage the SkillRepo.

Limitations and Open Questions

The paper does not detail the computational cost of RL training, a practical engineering concern.

Cross‑domain transfer claims lack quantitative bounds; more comparative experiments are needed.

While Markdown is convenient, the handling of skill dependencies, versioning, and conflicts remains unexplored.

Conclusion

The author reiterates that skills will become a standard component of agent systems, and SkillOS demonstrates that not only must agents use skills, they must also learn to manage them autonomously.