代理系統與工具路由
CAM-DF Reframes Agent Tool Ranking as a Cost-Aware Stopping Decision, Reducing Tool Exposure by 37%
New research argues that relevance ranking can determine tool order but cannot answer how many tools an agent should load. CAM-DF directly learns the marginal utility of continuing to load tools, reducing tool exposure by 37% while maintaining a comparable task success rate.
Tool routers typically compute relevance scores for search tools, databases, and connectors, then select a fixed top-k or truncate the list at a threshold. However, this approach does not jointly account for differences in latency, monetary cost, context usage, and data-exposure risk across tools. A study submitted on July 29 reframes the problem as a “cost-aware stopping decision over ranked prefixes” and introduces CAM-DF and the more interpretable CAM-DF-lite.
During training, the system calculates the utility difference, for each candidate prefix, between “stop now” and “continue to the ex post optimal position.” The sign of the difference becomes the stopping label, while its magnitude is used to assign greater weight to costly misclassifications. The model therefore learns not whether the current tool set appears sufficient, but whether the expected value of the next tool can offset its cost. The paper also proves that when tool costs are heterogeneous, monotonic truncation based solely on relevance scores generally cannot recover the optimal decision.
The study covers five tool-use domains and 1,343 tasks. On τ-bench Retail, CAM-DF delivers higher utility than a “predict sufficiency, then apply a threshold” baseline across five upstream rankers and two cost settings. Its advantage is more pronounced under greater cost pressure or with weaker rankers. In end-to-end execution, agents are exposed to 37% fewer tools than with full loading, while the observed task success rate remains broadly comparable. This is especially important for large MCP or enterprise connector catalogs: loading fewer tools not only saves tokens but also narrows the range of sensitive schemas, descriptions, and parameters visible to the model.
However, CAM-DF currently selects a prefix along a predetermined ranking before execution. It does not yet reassess the next step based on tool-returned content, and it depends on offline utility labels and manually specified cost functions. The paper does not include a standalone library, and some experiments reuse τ-bench; the benchmark’s original tasks were subsequently revised. Engineering teams should next evaluate whether the method can be reproduced on the updated tasks and under real-world latency and permission costs, and extend it into a sequential decision-maker that can update after every tool call.