Trace learning for self-improving agents

This post introduces Glean's approach to trace learning for self-improving AI agents. It details a two-component system: offline learning to mine strategies from historical traces and online application to retrieve and apply those strategies at runtime. The offline process involves a 'teacher' agent and multiple 'student' agents to score and distill effective strategies. The online application retrieves relevant memories to adapt agent plans before execution. The post emphasizes designing for enterprise accuracy by extracting assertions across multiple responses, verifying conflicts with Glean Search, and avoiding learning when inconsistencies arise. It also highlights the importance of storing intentionally narrow tool strategies that generalize across related tasks and splitting learnings into deployment and user levels for security and personalization.