Workflow-to-Skill: Skill Creation via Routing-Workflow-Semantics-Attachments Decomposition

Researchers introduce W2S, a framework for automatically constructing high-quality skills for large language model agents by decomposing execution traces into workflow structures, semantics, and attachments. The approach outperforms traditional summarization methods by 10.5%, demonstrating that treating traces as executable specifications rather than text yields more reliable agent behavior.

The paper addresses a critical bottleneck in scaling AI agent systems: the manual creation of procedural skills that encode domain-specific knowledge. As LLM agents become more prevalent in production environments, the cost of hand-writing quality skills limits deployment at scale. W2S tackles this by learning from diverse interaction evidence—demonstrations, trajectories, logs—automatically extracting reusable skill definitions that maintain behavioral consistency.

The technical insight distinguishing this work from prior approaches lies in rejecting simple text summarization. Agent traces contain fragmentation, redundancy, and edge cases that summarization loses. The RWSA intermediate representation captures three dimensions: workflow (task decomposition and control flow), semantics (verification and safety guarantees), and attachments (rollback and state management). This three-part decomposition directly mirrors production system requirements, making extracted skills immediately deployable rather than requiring manual refinement.

For the AI development community, this research accelerates the transition from hand-crafted to learned-from-evidence skill libraries. The 10.5% improvement in behavioral replay consistency indicates that the framework preserves safety-critical behaviors often missed by simpler approaches. This matters particularly for high-stakes domains where agents must handle rare but dangerous edge cases.

Looking forward, the framework's reliance on diverse trace sources suggests a path toward continuous skill improvement through production monitoring. As agents execute tasks, their traces feed back into the learning process, potentially creating a virtuous cycle of skill refinement. The question becomes whether this approach scales to complex, open-ended domains where trace coverage remains sparse.

• →W2S automatically constructs LLM agent skills from execution traces, eliminating costly manual creation.
• →The RWSA representation preserves safety-critical behaviors and control flow that text summarization loses.
• →Experimental results show 10.5% improvement in behavioral consistency over existing baselines.
• →The framework treats execution traces as runtime specifications rather than compressible text.
• →Three-part decomposition (workflow, semantics, attachments) directly supports production deployment requirements.