A preprint posted on arXiv on 9 October by Texas Tech University researchers introduces RATED, a method designed to help AI agents find tools matching a request without repeatedly selecting tools that are unavailable or failing. It first ranks candidates by semantic similarity, then updates the ranking from execution feedback. The researchers created EcoAgent for evaluation, a benchmark with 5,172 real-world tools and more than 210,000 query-tool pairs.

On the EcoAgent-Real test of real queries, the paper reports a 52.2% increase in mean reciprocal rank over the ToolLLM API retriever; RATED’s MRR of 0.618 remained below Ada-002 at 0.670 and SPLADE-v3 at 0.626. In a simulated scenario where 30% of tools failed, RATED reduced the fault-selection rate to 2.9%. This is a controlled benchmark and injected-failure test, not a reliability guarantee validated in live agent deployments; cold-start behavior on newly added tools is also left for future study. The work is a preprint that has not yet undergone peer review.