NVIDIA Nemotron Achieves Benchmark-Leading Performance With LangChain Deep Agents Harness
17:00 · July 8, 2026 · NVIDIA
NVIDIA Nemotron 3 Ultra is offering leading performance at lower cost than top closed models with the largest and most widely adopted AI agent orchestration platform. LangChain tuned its Deep Agents harness for NVIDIA Nemotron 3 Ultra, achieving the highest accuracy among open models, while completing more tasks at higher throughput and running at 10x […]
Summary
NVIDIA Nemotron 3 Ultra, paired with a tuned version of LangChain’s Deep Agents harness, reaches accuracy levels that match the top closed models on LangChain’s public benchmark while completing more tasks at higher throughput. The performance gains required no retraining of the underlying model. Instead, LangChain engineers examined execution traces and adjusted system prompts, tool descriptions and middleware to reduce points lost during agent runs. The result is inference cost reported at one-tenth that of comparable closed models, allowing teams to run continuous evaluations and iterate on specialized agents without proportional budget increases.
The arrangement supplies an open reference blueprint called NVIDIA NemoClaw for LangChain Deep Agents. It combines the tuned LangChain Deep Agents code with NVIDIA OpenShell, a secure runtime that executes agent actions inside enterprise environments. Because the model, orchestration layer and runtime are all open, organizations retain end-to-end control over data flows, customization and deployment location—on premises, in their chosen cloud or under internal governance policies.
Early adopters such as Abridge, Amdocs and Box are already embedding these agents into production platforms, while systems integrator EY is extending its NVIDIA practice to help clients adapt the blueprint for high-value workflows. Developers can obtain the tuned harness directly from LangChain or start from the NemoClaw blueprint; hosted access to Nemotron 3 Ultra is available on several inference platforms. The approach illustrates how environment-level engineering can deliver competitive agent performance while preserving ownership of the full stack.
Why it matters
This development is highly relevant as it offers a cost-effective, open-source alternative to closed AI models, which is crucial for driving AI adoption among Dutch SMEs. Furthermore, the ability to run these agents on proprietary infrastructure aligns perfectly with European data sovereignty and strict AI governance requirements.
