SkillTrace: A Graph-Based Approach for Composable LLM Agents

SkillTrace is a novel method for composing reusable skills in large language model (LLM) agents, introduced in a paper on arXiv (2608.02356). The approach addresses the challenge of identifying complete and executable skill compositions by organizing user queries into a semantic hierarchy, matching skill queries with candidates, and propagating over skill dependencies. This is achieved through a three-level graph that captures compositional relations among skill queries, similarity between queries and candidates, and dependencies among selected candidates. Experiments on SkillsBench and ALFWorld demonstrate state-of-the-art performance, with success rates of 53.17% and 91.43%, respectively. SkillTrace also shows consistent improvements across different backbone language models, indicating its robustness and generalizability. The paper argues that the problem of skill composition is best solved through this graph-based framework, which goes beyond simple retrieval of individually relevant skills. The method's effectiveness on two distinct benchmarks suggests its potential for enhancing LLM agents in complex task-solving scenarios.