Compositional Skill Routing for LLM Agents: Decompose, Retrieve, and Compose

The real contribution here is not retrieval itself but the dependency-aware composition step: most prior work treats tool selection as a flat ranking problem, and SkillWeaver's execution plan assembly is what makes multi-step automation tractable at the scale of thousands of real server skills rather than toy toolsets.

The inference efficiency angle running through recent coverage is worth connecting here, even if the link is indirect. The ConSA paper from the same day addresses how learned allocation patterns can replace static heuristics in attention routing. SkillWeaver is solving an analogous problem one level up: replacing hand-coded dispatch logic with learned embeddings for skill routing. Both papers are pushing against the same underlying pattern, which is that static, human-designed rules break down as model systems scale. The difference is that ConSA operates inside the model at inference time, while SkillWeaver operates above it at the agent orchestration layer. They are not competing approaches, but together they suggest a broader shift toward learned routing at multiple levels of the stack.

CompSkillBench is currently the only evaluation grounded in real MCP server skills at this scale, so watch whether other agent frameworks adopt it as a shared benchmark within the next six months. If they do not, the evaluation remains self-referential and the generalization claims stay unverified.