Debugging Deep Agents with LangSmith

Unlike simple LLM calls, deep agents perform tens to hundreds of steps over several minutes, involving long prompts and multi-turn interactions. This results in massive traces for a single run, making it difficult for humans to directly identify errors.

LangChain has enhanced the debugging capabilities of LangSmith to address these issues. The key updates are as follows:

• Polly: An AI assistant that helps with agent engineering by analyzing massive traces to support debugging.
• langsmith-fetch CLI: A command-line tool that grants debugging capabilities to coding agents such as Claude Code.

These tools track decision-making, prompt instructions, and tool-calling processes within agents, enabling rapid identification of error causes.