RATED teaches officers how to avoid tools that are unavailable when they need them

A suitable tool may be out of service. RATED uses execution feedback to reorder the options, based on an assessment of over 5,000 real-world tools.

RATED combines the search for tools by semantic proximity with a light mechanism that learns from their reliability during execution. The objective is to prevent an agent from continuing to select a failed API simply because its description corresponds to the demand.

The EcoAgent test bench consists of 5,172 tools and more than 210,000 pairs of queries and tools. In the high-unavailability scenarios evaluated, the authors report a decrease in the selection of failed tools from 86.2% to 2.9%, with less than a millisecond of additional rescheduling costs. These figures describe this experimental protocol.

Pre-publication filed on arXiv on 9 October 2026; The results are those reported by the authors and remain to be confirmed independently.