Tuning the harness, not the model: a Nemotron 3 Ultra playbook
The Nemotron 3 Ultra harness was tuned to match the model's capabilities, achieving a best run of 0.86 on the Deep Agents suite, nearly matching Opus 4.8's best of 0.87, at a lower cost. The harness was tuned through a trace-driven loop, ensuring validation across multiple trials without regression, and allowed the model to focus on the task rather than the scaffolding. A matched harness improves performance by optimizing the model's capability allocation, while a mismatched one forces the model to combat the scaffolding. The tuning process was data-driven, using evaluations to refine the harness and improve performance.
More from Langchain
Why managed agents are the next big thing in agent building
Managed Deep Agents gives developers a managed way to build, run, and deploy Deep Agents with built-in runtime, streaming, sandboxes, evals, memory, and auth.
langchain.com··Mads Kristian Nylund·AI InfrastructureAI-agents
LangSmith BYOC on AWS is generally available
LangSmith Bring Your Own Cloud is now generally available on AWS, giving Enterprise teams managed observability, evaluation, and deployment inside their own VPC.
langchain.com··Mads Kristian Nylund·AI DevelopmentAI InfrastructureAI Platforms
What is an AI agent?
Learn what AI agents are, how they work in an LLM loop, and where workflows fit so you can build reliable, production-ready autonomous systems.
langchain.com··Mads Kristian Nylund·AI DevelopmentAI InfrastructureAI-agents