Reading Skill Routing Out of a Frozen LLM's Own Forward Pass Beats Retrieval Pipelines That Add 16B Parameters

arXiv 2609.15982·medium signalXBlueskyLinkedInCopy link

Deployed harnesses route to skills by preloading every skill's metadata into context, which disperses attention and caps library size; retrieval pipelines move selection out of context but also out of the agent's capability. Gavel trains two linear maps to read mid-layer states, scoring the full library against per-skill banks built in one forward pass at installation, then resumes the shortlisted skills' forward passes for a yes/no verdict fused as a product of experts. On Qwen3-32B it beats progressive disclosure and retrieve-and-rerank pipelines that add 1.2B to 16B external parameters by up to 13.4 points on written tasks and 21.9 points when the need for a skill arises mid-rollout, with no skill text in context.