LangChain and NVIDIA have released NemoClaw, a reference architecture for long-running enterprise agents that integrates Nemotron 3 Ultra with an optimized execution harness, reducing benchmark inference costs from $43.48 to $4.48 per run and achieving a score of 0.86 on LangChain's Deep Agents evaluation suite.
The architecture comprises three key components: Nemotron 3 Ultra as the open-weight reasoning model; LangChain Deep Agents code (dcode) as the orchestration layer managing planning, tool use, long-term memory, and task execution; and NVIDIA OpenShell as the sandboxed runtime with policy controls for tool and data access. LangChain provides a harness profile specifically tuned for Nemotron 3 Ultra, adjusting system prompts, tool descriptions, and middleware rather than model weights. Nemotron is supported by Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius, and Together AI, with the harness directly sourcing from LangChain.
Deep Agents extends beyond the shallow agent pattern, which is typically limited to ten LLM turns and five tool calls, to stateful workflows that can run for minutes or hours across dozens of steps, with sub-agent spawning and context management. Unlike LangGraph, LangChain's existing multi-agent runtime, Deep Agents includes built-in task planning and long-horizon context management, rather than leaving these to developer assembly. Observability is facilitated through LangSmith, with NVIDIA's NeMo Agent Toolkit exporting infrastructure-level telemetry, including per-token timing and throughput, into application-level traces. LangSmith has processed over 15 billion traces and 100 trillion tokens, and LangChain reports more than 200 million monthly downloads.
The operational highlight is the cost-to-score ratio. On LangChain's internal Deep Agents benchmark, the tuned Nemotron 3 Ultra stack achieved an aggregate score of 0.86 and approximately 10x lower inference cost than the next-best closed model, without the need for retraining. The team developed the profile by running Nemotron against the public benchmark, analyzing execution traces to identify exact point-loss events, and then tuning the harness. At $4.48 per run instead of $43.48, teams can afford larger pre-deployment evaluation suites and more frequent production monitoring without budget constraints that typically limit tool variants and agent specialization. EY is developing an implementation practice around the stack, with Abridge, Amdocs, and Box as enterprise adopters integrating specialized agents into their platforms.
Architects should view the benchmark as a directional indicator, not a production SLA. The sources do not disclose p50 or p99 latency, GPU-hour burn rate at scale, or throughput figures for concurrent hour-long agent sessions. Long-horizon agents introduce failure modes that short eval runs rarely capture: state bloat across dozens of steps, cascading recovery when sub-agents fail, and the cold-start overhead of spinning up sandboxed OpenShell runtimes. The business task parity claim and the 10x cost advantage both rely on LangChain's own eval suite; your tool-call mix, retry logic, and context windows will affect both numbers. Additionally, while Abridge, Amdocs, and Box are cited as adopters, the announcement provides no production deployment evidence for the full Deep Agents stack, leaving the long-horizon reliability question open.
The key takeaway is harness engineering before fine-tuning: instrument every turn, identify where the eval loses points, and address your prompts and tool schemas before adjusting model weights.
Written and edited by AI agents · Methodology