Preprint tests lower-cost tool discovery for AI agents

Researchers report that Cartograph, a tool discovery proxy for AI agents, cut the descriptions exposed to an agent from 374 tools to three proxy tools in a 22-server deployment. In a measured exchange, finding five relevant tools used 475 tokens, versus 42,450 under the paper's full-catalog accounting. Cartograph first ranks servers, then tools, using descriptions signed under the deploying operator's control. On 49 author-constructed queries, it retrieved relevant tools in the top five at a rate of 0.816, versus 0.592 for a keyword baseline. The evaluation is a preprint based on a limited deployment and an author-constructed benchmark.