LangChain Deep Agents: Nemotron 3 Ultra harness tutorial

This tutorial examines how to improve an AI agent by changing the software around its model. A LangChain engineer uses Deep Agents with NVIDIA Nemotron 3 Ultra to demonstrate an evaluation, inspect a failure and register a custom harness profile without changing model weights.

The example asks the agent to return the last nonempty line of a 301-line file. In the initial run, it reads only part of the file and answers with X instead of the final marker. LangSmith traces expose the incomplete read. The engineer adds middleware that appends the total line count, remaining lines and next offset to a partial read result. In the demonstrated rerun, the agent continues reading and returns the marker.

The speakers recommend testing a targeted cluster of failures first, then running the broader evaluation suite and holdout cases. They describe how a change that helps file operations can hurt another category, so one passing example does not establish general reliability.

The discussion also distinguishes the general Deep Agents harness from coding-focused OpenCode and considers routing between local models and frontier models. It explains application-side harness tuning rather than providing hardware requirements or a step-by-step guide to run models locally.