When building AI agents that use external tools, tracking which tools actually work at runtime and adapting retrieval accordingly is as important as semantic matching—RATED shows this reduces failures from 86% to 3% during outages with minimal overhead.
RATED is a tool retrieval system for AI agents that combines semantic matching with real-time reliability feedback. Unlike existing methods that only consider relevance, RATED adapts to actual tool failures during execution using a lightweight bandit algorithm, enabling agents to avoid broken tools without retraining.