GoSkills: Group-Structured Skill Retrieval for Agent Skill Libraries
A new method called Group of Skills (GoSkills) has been developed by researchers to enhance skill-augmented agents through inference-time group-structured retrieval. Unlike current approaches that provide either standalone skills or bundles with implicit roles, which necessitate agent inference for execution points, support skills, requirements, and failure avoidance, GoSkills transforms the retrieval process. It shifts from a simple skill list to a concise, role-labeled execution context. By constructing anchor-centered skill groups from a typed skill graph and expanding support groups via a group graph, it narrows down the chosen group plan into a limited set of atomic skill payloads. This method establishes a fixed execution contract featuring Start, Support, Check, and Avoid fields, without needing modifications to the downstream agent, skill payloads, or execution. The research can be accessed on arXiv with the identifier 2605.06978.
Key facts
- GoSkills is an inference-time group-structured retrieval method for skill-augmented agents.
- Existing methods return atomic skills or dependency-aware bundles with implicit roles.
- GoSkills changes the retrieval object to a compact, role-labeled execution context.
- It builds anchor-centered skill groups from a typed skill graph.
- It expands support groups through a group graph.
- It bottlenecks the selected group plan into a bounded set of atomic skill payloads.
- It renders a fixed execution contract with Start, Support, Check, and Avoid fields.
- The method does not require changes to the downstream agent, skill payloads, or execution.
- The paper is available on arXiv under identifier 2605.06978.
Entities
Institutions
- arXiv