SkillGen: Multi-Agent Framework for Synthesizing Auditable LLM Agent Skills
A new multi-agent framework called SkillGen has been developed by researchers to create auditable skills for LLM agents, utilizing trajectories produced by a base agent. Unlike traditional skill creation methods, SkillGen employs contrastive induction to analyze both successful and unsuccessful trajectories, allowing it to pinpoint reusable success patterns, common failure modes, and behaviors that appear in successes but are missing in failures. The framework not only generates potential skills but also refines them through iteration. A significant innovation lies in modeling agent skills as interventions, which allows for empirical verification of their impact on overall performance through outcome comparisons. The resulting output is a human-readable document that can be reviewed prior to implementation, enhancing agent capabilities without the need for retraining while ensuring the process remains reusable and manageable. The paper can be accessed on arXiv with the identifier 2605.10999.
Key facts
- SkillGen synthesizes auditable skills from base agent trajectories.
- Uses contrastive induction over successful and failed trajectories.
- Identifies reusable success patterns, failure modes, and behavioral gaps.
- Generates and iteratively refines candidate skills.
- Models skills as interventions to verify net effect on performance.
- Output is a human-readable, inspectable artifact.
- Improves LLM agent capabilities without retraining.
- Paper available on arXiv: 2605.10999.
Entities
Institutions
- arXiv