Skills
Cap how many tools you register, not just which: cost-aware stopping cut live tool exposure 37% at comparable success across 1,343 tasks
CAM-DF reframes tool acquisition as a stopping problem — it trains on the offline gap between stopping now and the best continuation, using the sign to label the decision and the magnitude to weight each error by the payoff at stake, proving score-only relevance rules are suboptimal under heterogeneous tool costs. Across five tool-use domains and 1,343 tasks it reduced tool exposure 37% in live execution while holding task success. It ships as a lightweight pre-execution plugin over an existing tool ranking with no LLM fine-tuning, which makes it directly applicable to anyone whose MCP server list has quietly grown past what the context window should carry.
↳ Follow the thread