Generative Skill Composition

Analysis

Why This Paper Matters

As LLM agents increasingly rely on modular skills to tackle complex tasks, the challenge of selecting the right combination of skills becomes critical. Existing methods either expose the entire skill library to the agent's reasoning or use retrieval techniques, but both overlook the structural nature of skill composition—the joint decision over which skills, how many, and in what order. This paper identifies this gap and formalizes it as structured skill composition, a problem that cannot be solved by decoupling these dimensions. By proposing SkillComposer, the authors introduce a novel approach that treats skill composition as a sequence prediction task, enabling joint optimization of all three dimensions in a single decoding pass. This is a significant step forward because it directly addresses the combinatorial complexity of skill selection while maintaining efficiency.

The paper's relevance is underscored by the growing use of skill libraries in production coding agents. The empirical results show substantial gains over strong baselines, indicating that the structured approach captures dependencies that retrieval methods miss. This work is likely to influence future agent designs, encouraging more principled handling of procedural knowledge.

Technical Contributions

• Formalization of structured skill composition: The paper defines the problem as predicting an executable skill plan that jointly specifies the activated subset, count, and execution order, highlighting the interdependence of these dimensions.
• SkillComposer architecture: A constrained autoregressive decoder over skill identifiers that generates skill sequences. The constraints ensure valid plans, while the autoregressive nature captures dependencies between successive skills.
• Training data construction: The authors build a dataset of task-composition pairs from a real, human-curated skill library, enabling supervised training of the model.
• Efficiency: By generating the plan in a single pass, SkillComposer avoids the overhead of retrieving and reranking multiple candidate skills, reducing prompt-token cost.

Results

The evaluation on SkillsBench with two production-grade coding agents (GPT-5.2-Codex and Gemini-3-Pro-Preview) shows that SkillComposer raises the pass rate by +23.1 and +18.2 percentage points over the no-skill baseline, respectively. It surpasses top-3 retrieval methods and matches the gold-skill retrieval upper bound, which assumes perfect retrieval. Notably, this is achieved at a lower prompt-token cost, making it both more accurate and more efficient. These results demonstrate that jointly modeling skill composition yields better performance than decoupled retrieval or full exposure.

Significance

This work addresses a central bottleneck in LLM agent skill reuse, offering a scalable and efficient solution that could be applied across domains beyond coding. By formalizing the problem and providing a practical method, it opens avenues for future research in structured prediction for agent planning. The approach's efficiency also has practical implications for reducing API costs in production systems. As skill libraries grow, SkillComposer's ability to generate compact, executable plans will become increasingly valuable, potentially enabling agents to handle more complex tasks with limited context windows.