LangChain and NVIDIA launch NemoClaw, an open agent stack at a tenth of the cost
Nemotron 3 Ultra, the Deep Agents harness, and a governed runtime, with a benchmark run priced at $4.48 against $43.48 for the nearest closed rival.
- LangChain and NVIDIA launched the NemoClaw Deep Agents Blueprint on July 8, an open-source enterprise agent stack available now at build.nvidia.com.
- Three layers: NVIDIA's open Nemotron 3 Ultra model, LangChain's Deep Agents Code harness for planning, tool use and memory, and NVIDIA OpenShell, a governed sandboxed runtime.
- On LangChain's public Deep Agents benchmark, Nemotron 3 Ultra scored 0.86 aggregate at $4.48 per run, versus $43.48 for the next-closest model.
- NVIDIA frames it as roughly 10x lower inference cost than closed-model agent stacks.
The orchestration layer just made its strongest bid yet to capture the agent economy's margin. On July 8, LangChain and NVIDIA launched the NemoClaw Deep Agents Blueprint, a fully open enterprise agent stack: NVIDIA's Nemotron 3 Ultra as the model, LangChain's Deep Agents Code harness handling planning, tool use, memory and task execution, and NVIDIA OpenShell as a governed, sandboxed runtime that enforces policies on what agents can touch.
The number that sells it
On LangChain's public Deep Agents evaluation suite, Nemotron 3 Ultra posted a 0.86 aggregate score at $4.48 per benchmark run. The next-closest model cost $43.48 per run. Those figures come from LangChain's own benchmark, so treat them as a vendor's home field, but the roughly 10x cost gap is the whole argument: NVIDIA's framing is that enterprises can run serious agents on an open stack without paying closed-model prices. Jensen Huang went further in the release: "Super agents have arrived... companies will use AI cloud services and build their own AI, shaped by their proprietary data, know-how, and workflows."
Tune the harness, not the model
The more interesting claim is architectural. LangChain CEO Harrison Chase argues the differentiation has moved out of the model and into the system wrapped around it: "The way to build better agents is to keep improving the system around the model. Memory, tool use, evaluation and model behavior compound when teams can tune them together." That is a direct play for value against the frontier labs, and it explains why the blueprint ships with an ecosystem already attached: EY, Baseten, Fireworks, Nebius, Crusoe, DeepInfra and Together AI are all partners at launch.
What it means for the stack wars
Three months ago Microsoft unified its agent frameworks to own the enterprise harness. Now the open ecosystem has a counterpart with NVIDIA's weight behind it, and a cost story that lands directly on agent unit economics. Every vendor selling outcomes at a fixed price improves its margin the moment a run costs $4 instead of $40. The same week that frontier model pricing split in two, the orchestration layer started selling the argument that where the intelligence comes from matters less than what harness it runs in.
Sources: LangChain, NVIDIA, PRNewswire.