Trace learning for self-improving agents

What this post added

This post details Glean's approach to trace learning for self-improving AI agents. It introduces the concept of traces as records of task execution and explains how trace learning enables agents to learn from successes and failures to improve future performance. The post outlines a two-component system: offline learning, which mines strategies from historical traces using a teacher-student agent comparison, and online application, which retrieves and applies learned strategies at runtime. It emphasizes the importance of accuracy in enterprise environments, detailing a stricter validation process involving cross-response assertion checks and conflict verification. The post also highlights the design choice to store intentionally narrow tool strategies that generalize across related tasks, focusing on workflow-level patterns rather than individual tool calls. Finally, it explains how learnings are split into deployment and user levels for secure capture of individual preferences, with actions replayed in a sandbox for safety.