NVIDIA NeMo Agent Toolkit

Scroll sideways for more →

The NVIDIA NeMo Agent Toolkit is an open-source, framework-agnostic library for connecting, profiling, evaluating, and optimizing teams of AI agents. Released by NVIDIA in 2025, first as AgentIQ and briefly as the Agent Intelligence Toolkit (AIQ Toolkit), it adds a common operational layer (instrumentation, observability, evaluation, optimization, and reusable connectors) on top of agentic AI systems without forcing teams to rebuild them on a new platform.^[1][7] It is distributed as the PyPI package nvidia-nat and belongs to the broader NVIDIA NeMo software family.^[2][4] At SIGGRAPH 2026 the toolkit was expanded with new Omniverse libraries that let agents build simulation-ready 3D worlds.^[11]

What is the NeMo Agent Toolkit

The NeMo Agent Toolkit is not a new agent framework meant to replace the ones developers already use. It is a layer that sits alongside them. NVIDIA describes it as an open-source library "for efficiently connecting and optimizing teams of AI agents," and, in its own documentation, as a way to add "intelligence to AI agents across any framework, enhancing speed, accuracy, and decision-making through enterprise-grade instrumentation, observability, and continuous learning."^[1][3]

The central design idea is that every agent, tool, and multi-step workflow is treated as a composable function call. Because the unit of composition is a plain function call rather than a framework-specific object, components written for different agent stacks can be mixed, wrapped, and reused inside one system.^[1][7] On top of that abstraction the toolkit layers the operational tooling that production agent systems usually lack: end-to-end profiling (down to individual tokens), observability, a repeatable evaluation harness, and automated optimization.^[1][7]

In practice this means a team that has already built a multi-agent application (for example a LangGraph research agent calling a CrewAI crew and some custom Python tools) can wrap those pieces with the toolkit to measure where time and tokens are spent, trace execution, run evaluations, and then tune the workflow for speed, accuracy, and cost, all without replatforming.^[1][9] Early coverage described it as "a lightweight and flexible Python library designed to unify agentic workflows across frameworks, memory systems, and data sources," functioning "as a complementary layer rather than a replacement framework."^[7]

Name history (AgentIQ, Agent Intelligence Toolkit, NeMo Agent Toolkit)

The project has been renamed twice, but NVIDIA has kept the codebase, API, and version numbering continuous across each change, so upgrading has mostly meant changing a dependency name.^[2][6] It began life as AgentIQ, introduced alongside the open Llama Nemotron reasoning models at NVIDIA's GTC conference in March 2025 and open-sourced in early April 2025.^[8][7] It was then renamed the Agent Intelligence Toolkit (shortened to AIQ Toolkit), and finally the NeMo Agent Toolkit, a change NVIDIA said was made "to better reflect the purpose of the toolkit and to align with the NVIDIA NeMo family of products," with "the core technologies, performance and roadmap" unchanged.^[6]

The package renames track the version history recorded in the toolkit's own migration guide:^[2]

Scroll sideways for more →

The acronym NAT in the package name stands for NeMo Agent Toolkit.^[1] NVIDIA advised existing users to update their dependency from agentiq or aiqtoolkit to nvidia-nat; compatibility aliases were kept for a few releases and then removed at v1.2.0 and v1.5.0 respectively.^[2] The current stable release, v1.8.0, shipped on June 16, 2026 under the Apache 2.0 license and supports Python 3.11, 3.12, and 3.13.^[4][5]

Architecture and how it works

The toolkit is organized around three activities: building agents, gaining insight into them, and optimizing them.^[1][3]

Framework-agnostic wrapping. The toolkit connects to existing agent frameworks rather than owning the agent logic. It integrates with LangChain and LangGraph, LlamaIndex, CrewAI, Microsoft Semantic Kernel, Google ADK, and Agno, as well as custom enterprise frameworks and plain Python agents.^[1] Because agents and tools are exposed as function calls, the same tool can be reused across frameworks, and reusable connectors let agents reach data sources and services without bespoke glue code.^[1][7]

Profiling. The toolkit instruments a workflow end to end, tracking token usage, response timings, and hidden latencies at a fine grain. NVIDIA describes the ability to "profile entire workflows from the agent level all the way down to individual tokens," which surfaces the bottlenecks that make multi-agent systems slow or expensive.^[1][7]

Observability. Telemetry is emitted in an OpenTelemetry-compatible form, so traces and metrics can flow into existing observability platforms (for example LangSmith or other tracing backends) to show how each part of a workflow behaves in production.^[1][7]

Evaluation. A built-in evaluation system provides a consistent, repeatable way to validate agentic workflows, including retrieval-augmented generation (RAG) pipelines and full end-to-end runs, so accuracy can be checked as a workflow evolves.^[1][7]

Optimization. Beyond measurement, the toolkit can tune workflows: hyper-parameter and prompt optimization, reinforcement-learning fine-tuning, integration with NVIDIA Dynamo for inference efficiency, and a set of Agent Performance Primitives (APP) that support parallel execution and priority routing across a multi-agent workflow.^[1] A built-in chat UI lets developers interact with a wrapped workflow while these tools run.^[1]

Interoperability protocols. The toolkit supports the Model Context Protocol (MCP) on both sides: it can act as an MCP client to connect to remote MCP servers, and as an MCP server to publish its own tools to any MCP client.^[12] It also works with Agent-to-Agent (A2A) workflows, letting agents built with the toolkit cooperate with agents running elsewhere.^[1][12] This is the same interoperability layer that patterns such as deep agents rely on when they orchestrate long-running, multi-step work.

Framework integrations

The toolkit is deliberately vendor-neutral about which agent framework a team starts from. The frameworks and protocols documented as first-class integrations include:^[1][3]

Scroll sideways for more →

NVIDIA credits CrewAI, LangChain, LlamaIndex, and Semantic Kernel in the project's acknowledgments as open-source projects the toolkit builds on, and the documentation also lists further connectors and integrations beyond the core frameworks.^[4][3] The consistent message is that adoption should be additive: teams keep their framework of choice and gain the operational layer.^[1][9]

The SIGGRAPH 2026 expansion: Omniverse libraries and the local agent stack

At SIGGRAPH 2026, in an announcement dated July 20, 2026, NVIDIA expanded the Agent Toolkit with new NVIDIA Omniverse libraries that give AI agents tools to bring physical AI capabilities into existing 3D applications.^[11][12] The stated goal is to let agents build and validate "simulation-ready worlds," which NVIDIA frames as the environments where robots, factories, and autonomous systems are trained and tested before they are deployed.^[11] Jensen Huang, NVIDIA's founder and CEO, said: "The physical AI era will be built in simulation first. NVIDIA Agent Toolkit with Omniverse libraries brings AI agents into the 3D tools developers already use, helping build the simulation-ready worlds where robots, factories and autonomous systems are trained and tested long before they reach the real world."^[11]

Three new open-source Omniverse libraries, published on GitHub, extend agent skills into simulation and 3D asset work:^[11][15]

Scroll sideways for more →

NVIDIA demonstrated the libraries through "SimReady" Blender, described as "a sample workflow built in Blender with NVIDIA Omniverse libraries and NVIDIA NemoClaw," which brings agent-ready RTX sensor simulation, physics, and asset validation into Blender without pulling artists out of their existing environment.^[11][15] NVIDIA also announced a blueprint for integrating the Omniverse libraries into Blender, giving NemoClaw agents callable sensor-simulation and physics tools.^[12] Early adopters named in the release include SideFX, which is exploring how agents can support Houdini workflows using OpenUSD together with ovrtx and ovphysx, and PTC, whose Onshape platform is using OpenUSD and ovrtx to connect cloud design work with simulation; startups Palatial, Lightwheel, ForgeCAD, and Moonlake AI were also cited.^[11][15]

The companion SIGGRAPH blog described how the toolkit turns a workstation into a self-contained agent system. In a section on running personal AI agents locally on the NVIDIA DGX Station, NVIDIA said the Agent Toolkit brings together NVIDIA NemoClaw, the NVIDIA Nemotron 3 Ultra open model (which NVIDIA describes as a frontier 550-billion-parameter model optimized for DGX Station GB300 systems), the Omniverse libraries as agent-accessible tools and skills, and a secure runtime called NVIDIA OpenShell, in a single local system that needs no internet connection.^[12] NVIDIA characterizes OpenShell as "the open source secure runtime" that "keeps agents sandboxed and governed according to defined policies."^[12] The same setup exposes MCP client and server support so local agents can reach external tools and publish their own.^[12] According to NVIDIA, standing this environment up on a DGX Station takes three steps and roughly 30 minutes.^[12] More broadly, NVIDIA says the Omniverse-library workflows can run locally on a range of hardware, "from compact RTX-powered systems with NVIDIA RTX Spark to NVIDIA GB300-powered systems with NVIDIA DGX Station."^[15]

Relationship to NeMo and the rest of NVIDIA's agent stack

The 2025 rename folded the toolkit into the NVIDIA NeMo family, the company's broader suite for building, customizing, and operating AI models and agents.^[6] Within that stack the toolkit occupies a specific slot, and it is easy to confuse with the pieces around it:

• NeMo Agent Toolkit (this page) is the framework-agnostic operational layer: it wraps, profiles, evaluates, and optimizes whatever agents you already have.^[1]
• NVIDIA NemoClaw is an agent blueprint, not a wrapper. It packages a model, an agent harness, and a runtime together so developers can build custom autonomous agents; the toolkit is the surrounding framework that runs and optimizes agents like these.^[12]
• Nemotron 3 and other Nemotron models are the reasoning models that can back an agent; the toolkit is model-agnostic and can drive them through NVIDIA NIM or other endpoints.^[12][9]
• NVIDIA NIM microservices provide the packaged model inference the agents call; the toolkit and NIM were co-integrated into Microsoft Azure AI Foundry.^[9]

The toolkit is also shipped as part of NVIDIA AI Enterprise, positioning it as the connective tissue between NVIDIA's models, inference microservices, and, since SIGGRAPH 2026, its Omniverse simulation libraries.^[1][11]

Adoption and use cases

Because the toolkit is additive rather than a replacement, its adoption has largely come through integrations with existing platforms. When AgentIQ launched at GTC 2025, NVIDIA and Microsoft integrated it, together with NVIDIA NIM microservices, into Microsoft Azure AI Foundry; Microsoft framed the combination as a way to speed up enterprise agent projects that can otherwise run 9 to 12 months from start to production.^[9][10] It is also distributed inside NVIDIA AI Enterprise, which is how many organizations encounter it.^[1]

Typical use cases fall into a few buckets:

• Connecting and reusing agents and tools across frameworks so an enterprise does not have to standardize on one agent library.^[1][7]
• Profiling and cost control for multi-agent systems, identifying the tokens, calls, and latencies that dominate a workflow's time and spend.^[1][7]
• Evaluation and continuous improvement, running repeatable checks on RAG and end-to-end agent workflows as they change.^[7]
• Simulation and 3D content, the SIGGRAPH 2026 direction, where agents help build and validate SimReady worlds inside tools such as Blender, Houdini (SideFX), and PTC Onshape.^[11][15]
• Local, private agents assembled on a DGX Station with NemoClaw, Nemotron 3 Ultra, the Omniverse libraries, and the OpenShell runtime, for teams that want an air-gapped setup.^[12]

Significance

The toolkit's significance is less about any single feature and more about where it sits. The 2025 to 2026 wave of agentic AI produced many competing frameworks (LangChain and LangGraph, CrewAI, LlamaIndex, Semantic Kernel, Google ADK, and others), and enterprises frequently ended up with agents spread across several of them. By treating agents and tools as function calls and adding a shared layer for profiling, observability, evaluation, and optimization, the toolkit offers an interoperability and operations tier that does not force a team to pick one framework or rebuild.^[1][7] That "wrap what you have" posture, plus an Apache 2.0 license and OpenTelemetry-compatible telemetry, made it an easy addition to existing stacks.^[1]

The SIGGRAPH 2026 expansion is significant for a different reason: it extends the same agent layer from text and tool use into physical AI and simulation, giving agents callable sensor-simulation and physics skills inside the 3D applications that artists and engineers already use, and packaging a full local agent runtime for the DGX Station.^[11][12]

Limitations

• Open source, but tuned for NVIDIA silicon. The libraries and toolkit are Apache-licensed and free on GitHub, yet they are optimized for NVIDIA hardware; commentators noted that "open source that runs optimally on NVIDIA silicon is a particular kind of open source," and that extracting comparable value on competing hardware is harder.^[13]
• An operational layer, not a full agent framework. The toolkit assumes you already have agents built somewhere; it wraps and optimizes them rather than providing the agent logic itself, so it is a complement to, not a substitute for, an underlying framework.^[7]
• Governance of agent-produced work. Independent analysis has raised the question of who governs the code and outputs that toolkit-optimized agents generate, a concern that grows as agents take on more autonomous engineering tasks.^[14]
• Newer capabilities depend on specific hardware and are still rolling out. The DGX Station local stack targets GB300-class systems, and NVIDIA said the compact RTX Spark systems for the smaller end of the range would ship in the fall of 2026, so some of the SIGGRAPH capabilities were previews at announcement time.^[12][15]
• Vendor benchmark claims. Figures such as the Nemotron 3 Ultra parameter count, the "three steps, 30 minutes" DGX Station setup, and framework speedups are NVIDIA's own statements and should be read as vendor claims rather than independently verified results.^[12]

In plain terms (ELI5)

Imagine several AI assistants working together on a job. One is good at research, one at writing, one at pulling data, and each was built with different tools, almost like workers who each speak a different language. The NeMo Agent Toolkit is like a universal adapter plus a stopwatch plus a supervisor. The adapter lets the assistants plug into one another even though they were built differently. The stopwatch measures exactly how long each step takes and how much it costs, so you can find the slow, expensive parts. The supervisor then helps them work faster and more accurately. The important part is that nobody has to quit their job and start over: the toolkit wraps around the team you already have. The 2026 SIGGRAPH update added a new skill set so these assistants can also build and test realistic 3D worlds, the kind used to train robots and self-driving cars before they are let loose in the real world.

See also

• NVIDIA NemoClaw
• NVIDIA NeMo
• Nemotron 3
• NVIDIA Omniverse
• Model Context Protocol
• Agentic AI
• LangGraph
• NVIDIA DGX Station

References

1. NVIDIA, "NeMo-Agent-Toolkit (README)," GitHub. https://github.com/NVIDIA/NeMo-Agent-Toolkit ↩
2. NVIDIA, "Migration Guide," NeMo Agent Toolkit documentation, GitHub. https://github.com/NVIDIA/NeMo-Agent-Toolkit/blob/develop/docs/source/resources/migration-guide.md ↩
3. NVIDIA, "NVIDIA NeMo Agent Toolkit Overview," NVIDIA Developer. https://developer.nvidia.com/nemo-agent-toolkit ↩
4. "nvidia-nat," Python Package Index (PyPI). https://pypi.org/project/nvidia-nat/ ↩
5. NVIDIA, "Releases, NVIDIA/NeMo-Agent-Toolkit," GitHub. https://github.com/NVIDIA/NeMo-Agent-Toolkit/releases ↩
6. NVIDIA, "NVIDIA Agent Intelligence toolkit is now the NVIDIA NeMo Agent toolkit," NVIDIA Developer Forums, 2025. https://forums.developer.nvidia.com/t/nvidia-agent-intelligence-toolkit-is-now-the-nvidia-nemo-agent-toolkit/335881 ↩
7. MarkTechPost, "NVIDIA AI Released AgentIQ: An Open-Source Library for Efficiently Connecting and Optimizing Teams of AI Agents," April 5, 2025. https://www.marktechpost.com/2025/04/05/nvidia-ai-released-agentiq-an-open-source-library-for-efficiently-connecting-and-optimizing-teams-of-ai-agents/ ↩
8. NVIDIA, "NVIDIA Launches Family of Open Reasoning AI Models for Developers and Enterprises to Build Agentic AI Platforms," NVIDIA Newsroom, March 18, 2025. https://nvidianews.nvidia.com/news/nvidia-launches-family-of-open-reasoning-ai-models-for-developers-and-enterprises-to-build-agentic-ai-platforms ↩
9. Microsoft, "Accelerating agentic workflows with Azure AI Foundry, NVIDIA NIM, and NVIDIA AgentIQ," Microsoft Azure Blog, 2025. https://azure.microsoft.com/en-us/blog/accelerating-agentic-workflows-with-azure-ai-foundry-nvidia-nim-and-nvidia-agentiq/ ↩
10. InfoQ, "Azure AI Foundry Integrates NVIDIA NIM and AgentIQ," March 2025. https://www.infoq.com/news/2025/03/azure-foundry-nim-agent-iq/ ↩
11. NVIDIA, "NVIDIA Agent Toolkit Expands With New Omniverse Libraries, Putting AI Agents to Work Building Simulation-Ready Worlds," NVIDIA Newsroom, July 20, 2026. https://nvidianews.nvidia.com/news/nvidia-agent-toolkit-expands-with-new-omniverse-libraries-putting-ai-agents-to-work-building-simulation-ready-worlds ↩
12. NVIDIA, "At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI," NVIDIA Blog, July 20, 2026. https://blogs.nvidia.com/blog/siggraph-news-2026/ ↩
13. UC Today, "NVIDIA Omniverse Libraries: Open Source, But There's a Catch," July 2026. https://www.uctoday.com/immersive-workplace-xr-tech/nvidia-omniverse-libraries-agent-toolkit-siggraph-2026/ ↩
14. Mneme HQ, "NVIDIA NeMo Agent Toolkit Optimizes AI Agent Workflows. Who Governs the Code They Produce?" https://mnemehq.com/insights/nvidia-nemo-agent-toolkit-engineering-governance/ ↩
15. NVIDIA, "NVIDIA Agent Toolkit Expands With New Omniverse Libraries" (press release full text), GlobeNewswire, July 20, 2026. https://www.globenewswire.com/news-release/2026/07/20/3329895/0/en/NVIDIA-Agent-Toolkit-Expands-With-New-Omniverse-Libraries-Putting-AI-Agents-to-Work-Building-Simulation-Ready-Worlds.html ↩
16. NVIDIA, "NVIDIA NeMo Agent Toolkit Overview (1.8)," docs.nvidia.com. https://docs.nvidia.com/nemo/agent-toolkit/1.8/index.html