Researchers propose SkillComposer to address the bottleneck of selecting appropriate skill compositions for LLM agents by formalizing it as structured skill composition. This approach jointly predicts the activated subset, count, and execution order of skills through task-conditioned skill sequence prediction.
- SkillComposer uses a constrained autoregressive decoder over skill identifiers to capture dependencies between successive skills in a single pass.
- The model is trained on task-composition pairs derived from a real, human-curated skill library.
- Evaluations on SkillsBench show SkillComposer raises the pass rate by +23.1pp and +18.2pp over the no-skill baseline on GPT-5.2-Codex and Gemini-3-Pro-Preview respectively.
- The method surpasses top-3 retrieval performance while matching the gold-skill retrieval upper bound at a lower prompt-token cost.
This approach improves downstream task success for production-grade coding agents by providing a more effective way to compose skills from growing libraries.