Nvidia unveiled its Agent Toolkit at GTC on March 16, giving enterprises an open-source software stack for building, securing, and deploying AI agents tailored to specific business domains.
The toolkit launched with more than 17 corporate partners already on board, including Adobe, SAP, and Salesforce.
What’s actually in the box
The core pieces include the NeMo library for observability, which lets companies monitor what their agents are doing and why. There’s Nvidia OpenShell, a secure runtime environment that enforces policy-based execution. Rounding out the stack are Nemotron open models, Nvidia’s own reasoning models that power the agents’ decision-making capabilities.
The toolkit also integrates with existing AI orchestration frameworks like LangChain and CrewAI. Rather than forcing enterprises to rip out their current infrastructure, Nvidia designed the toolkit to slot into what’s already there.
