We expanded the NVIDIA Agent Toolkit to help developers build agents that use company data to answer questions and make predictions, with controls over what agents can access and do. Watch how Lila Sciences , Siemens and Dassault Systèmes are using AI agents to design potential therapies, plan factory... production and refine vehicle designs. See what’s new →show more

@LilaSciences @Siemens @Dassault3DS agents anstheyring with company data also means anstheyring about competitors. they track how they rank brands:

@LilaSciences @Siemens @Dassault3DS Controls over what they can access is the main thing for me. Agent surfacing company data to the wrong person is every company's nightmare

@LilaSciences @Siemens @Dassault3DS NVIDIA Agent Toolkit builds enterprise agents

@LilaSciences @Siemens @Dassault3DS Siemens letting agents plan factory production now, my spreadsheet feels very threatened

@LilaSciences @Siemens @Dassault3DS That's awesome! It's great to see how these companies are using AI to push boundaries in design.

@LilaSciences @Siemens @Dassault3DS The dividing line will be consequence. Reading company data to answer a question is one risk class, while acting on a prediction in a factory or a lab is another. Approval thresholds should rise with that consequence, not stay flat across tools.

@LilaSciences @Siemens @Dassault3DS the access controls are the part I'd want to see more of. can you scope an agent to read only on some data and get a log of every query it ran? that's usually what security teams ask first

@LilaSciences @Siemens @Dassault3DS The access controls matter more than another flashy demo. 🔒

@LilaSciences @Siemens @Dassault3DS I’m glad you’re giving developers control over what agents can read and what they can do 🩷

@LilaSciences @Siemens @Dassault3DS For the access controls, can you publish a test where an agent finds the right answer in a document the requesting user cannot read? The useful score is how often it refuses that answer, not just how often it retrieves the right one.

@LilaSciences @Siemens @Dassault3DS every enterprise agent demo: 'it answers questions from your company data,' then it stops exactly where the data governance question starts.

@LilaSciences @Siemens @Dassault3DS scoped access holds until a narrow-scope worker writes a note that a wider-scope worker reads and obeys

@LilaSciences @Siemens @Dassault3DS can u trace each prediction back to the table it used??

@LilaSciences @Siemens @Dassault3DS The access controls are the real product here, the demos are just the trailer

@LilaSciences @Siemens @Dassault3DS I'd want both halves. Build-time least privilege limits what the agent can reach, and a check before each tool call limits what it does there. OWASP's Agent Control Standard puts that check in a separate Guardian:

@LilaSciences @Siemens @Dassault3DS Access controls are the real unlock here. Building the agent is the easy week. Getting approval for what data it can touch is the hard quarter.