NVIDIA used its GTC Taipei keynote on 1 June 2026 to lay out the next phase of the enterprise AI agent stack. The announcements were less about new models for their own sake and more about how organisations actually run, secure and govern long-running autonomous agents at scale.
The headline pieces were the NVIDIA Agent Toolkit, a new Nemotron 3 Ultra open model tuned for long-running agents, deeper integration of the OpenShell secure runtime with Microsoft, Canonical and Red Hat, and partnership work with the likes of Cadence, Dassault Systèmes, Siemens, Synopsys, CrowdStrike and Palantir. Taken together, they describe an enterprise agent platform with explicit attention to harnesses, runtime policy, identity and skills.
What was actually announced
NVIDIA framed the Agent Toolkit as open source foundations for building secure, long-running agents that behave like digital coworkers. The toolkit pulls together Nemotron open models, NemoClaw blueprints, the OpenShell secure runtime and CUDA-X libraries that agents can now call as domain-specific skills.
Alongside the toolkit, NVIDIA unveiled Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model post-trained for orchestration frameworks including Hermes Agent, LangChain Deep Agents, OpenClaw, OpenHands and OpenCode. NVIDIA cites up to five times faster inference and up to 30 percent lower cost compared with open frontier models in its class for complex agentic workloads.
On the runtime side, NVIDIA and Microsoft are working on new Windows security primitives plus OpenShell to deliver identity, containment and policy controls for agents running natively on Windows. Canonical is integrating OpenShell with Ubuntu through supported snaps and OCI-compliant containers, and Red Hat is integrating OpenShell into its AI platform and contributing upstream to standardise how agents are managed on enterprise infrastructure. SAP and ServiceNow already embed OpenShell into Joule Studio and Project Arc respectively.
Why this matters for enterprise buyers
For enterprise buyers, the important shift is not the model count. It is the formalisation of an agent stack that mirrors the way mature enterprise software has always been bought and operated.
Until recently, most agent conversations focused on capability demos: can the model use tools, plan, retry, summarise, write code. Buyers are now asking different questions. Where does this agent run. Who owns it. What identity does it carry. What can it call. How is its behaviour logged. How is it stopped. How are its costs attributed. NVIDIA is responding by positioning OpenShell as the runtime where those controls live, and the Agent Toolkit as the building blocks that connect models, harnesses, blueprints and skills around it.
The choice of partners reinforces the message. Cadence, Dassault Systèmes, Siemens and Synopsys are not building consumer copilots. They are introducing autonomous AI engineers into regulated, high-value engineering workflows where verification, traceability and policy matter. CrowdStrike and Palantir are using Nemotron models to power long-running agents inside cybersecurity and operational decision-making, areas where unbounded autonomy is unacceptable.
The harness, runtime and skills pattern
NVIDIA is unusually clear in its language about the layers involved. An agent starts with a model. A harness turns the model into an agent by adding orchestration, context, memory, tool use and security. A runtime then sets policy and privacy controls around that agent so it can be deployed at scale. Skills give the agent access to domain capabilities, in this case CUDA-X libraries such as cuDF, cuOpt, AI-Q, NeMo, PhysicsNeMo and CUDA-Q.
That separation matters. It allows buyers to evaluate each layer independently: which models are acceptable, which harness fits an existing engineering culture, which runtime satisfies security and compliance, which skills make the agent genuinely useful in a specific business context. It also gives suppliers a clearer story to tell about where they fit and what they integrate with.
Implications for the wider agentic AI market
Three threads stand out for enterprise teams watching the space.
Runtimes are becoming a procurement category. OpenShell joins a growing list of control points, including AI gateways, identity platforms and policy engines. Enterprises buying agent platforms in 2026 should expect to evaluate the runtime as carefully as the model.
Open models are getting serious about long-running work. Nemotron 3 Ultra is targeted explicitly at long-running agents and is post-trained for popular harnesses. That signals confidence that production agents will not be a single API call but persistent, tool-using workers operating across hours, days and workflows.
Vertical agent products will accelerate. When platform vendors hand specialised teams a credible runtime and a tuned model, the bottleneck shifts to domain expertise. Expect more vertical agent launches across engineering, healthcare, security and operations in the months ahead.
What buyers should ask now
- Which runtime will our enterprise agents execute on, and what identity, policy and audit controls does it enforce?
- Are the models we depend on tuned for long-running, tool-using agent workloads, not only chat?
- How do harness, runtime and skills fit together in our current platform, and can we change any layer without rebuilding?
- Where do agent costs surface, and can we attribute them to teams, products and outcomes?
- Which production behaviours can security teams observe and stop in real time?
The Agentic Expo angle
Agentic Expo exists for buyers and suppliers who are past the demo stage. The GTC Taipei announcements show what the enterprise side of the market now expects: clear separation of model, harness, runtime and skills, explicit attention to identity and policy, and credible partners taking long-running agents into real production environments.
Suppliers that can describe where they sit in this stack, how they integrate with established control layers, and what evidence they have from real deployments will find a more attentive audience. Buyers walking the floor at Olympia in March 2027 will be asking exactly these questions.