Teams building agents already have logs and traces. The hard part is what comes next: measuring quality, writing evals, and pushing improvements.

This talk introduces LangSmith's new feature built on LangChain's Deep Agents harness: it proposes evaluators, surfaces issues on a board, and opens PRs. Under the hood, it uses memory, skills, and filesystem tools to close the loop.

This session walks through the architecture and demos the cycle from trace to fix.