Generative Skill Composition for LLM Agents

2026-06-30

AI Summary

This work addresses the challenge that large language model agents face in effectively coordinating the selection, quantity, and execution order of skills for complex tasks. It formulates skill composition as a structured sequence prediction problem and introduces a task-conditioned constrained autoregressive decoder that generates, in a single pass, a complete executable plan specifying the skill subset, repetition counts, and execution sequence—naturally capturing inter-skill dependencies. Trained on synthetically constructed task-skill pairing data, the method achieves substantial performance gains on the SkillsBench benchmark, improving by 23.1 and 18.2 percentage points over skill-free baselines on GPT-5.2-Codex and Gemini-3-Pro-Preview, respectively, while matching the performance of an oracle skill retriever and reducing prompting overhead.

Abstract

Recent LLM agents benefit from skills for solving complex tasks. Skills encapsulate modular packages of procedural knowledge and instructions for performing specialized tasks, such as setting up a sandboxed environment, running a test suite, or refactoring a function across multiple files. As skill libraries grow and become reusable across tasks and domains, selecting an appropriate skill composition has emerged as a central bottleneck. Existing approaches fall into two categories. One exposes the agent's reasoning to the entire skill collection; the other performs skill retrieval via embeddings or LLM-based rerankers. Both provide useful insights; however, they miss the structural nature of skill composition, which is a joint decision over which skills, how many, and in what order -- three dimensions that cannot be decoupled. We formalize this as structured skill composition: given a task and a skill library, predict an executable skill plan that jointly specifies the activated subset, count, and execution order. We propose SkillComposer, which instantiates structured skill composition as task-conditioned skill sequence prediction. SkillComposer uses a constrained autoregressive decoder over skill identifiers, so subset, count, and order emerge jointly from a single decoding pass, and dependencies between successive skills are captured naturally. We build a training set of task-composition pairs from a real, human-curated skill library. We then evaluate SkillComposer along two axes: composition quality on a held-out test set, and downstream task success on SkillsBench across two production-grade coding agents. On GPT-5.2-Codex, Gemini-3-Pro-Preview, SkillComposer raises the pass rate by +23.1, +18.2pp over the no-skill baseline, surpassing top-3 retrieval and matching the gold-skill retrieval upper bound at lower prompt-token cost.

Problem

• skill composition
• LLM agents
• structured prediction
• executable skill plan
• skill library

Innovation

• structured skill composition
• SkillComposer
• autoregressive decoding
• LLM agents
• skill planning