NVIDIA's Jacob Liberman on Agentic AI in the Enterprise
Recorded live at GTC 2025, NVIDIA Director of Product Management Jacob Liberman frames AI agents as the third era of GenAI, following chat/copilot usage and RAG. He explains that an agent is essentially an instruction-tuned LLM trained to use tools, likening it to giving a generalist model a vocational degree. Liberman argues that within a few years the majority of LLM tokens will be generated for agent-to-agent communication rather than human interaction, mirroring the rise of machine-to-machine trading in computational finance. On adoption challenges, Liberman highlights the lack of standardization across agent implementations, which creates friction when agents need to communicate and share memory. He stresses that enterprises require deterministic outcomes, and that agentic workloads can produce unpredictable token costs due to unbounded reasoning. He also notes a shift in enterprise meetings, where legal and HR now join IT and security to discuss governance. Liberman details NVIDIA's response: reference blueprints for AI agents (digital humans, biomedical research, cybersecurity vulnerability analysis, and simulation-to-real-world robotics) available at build.nvidia.com/blueprints and on GitHub. He describes reasoning models and test-time compute as the dominant enterprise paradigm, with use cases spanning software co-development, document summarization, and rewriting previously solved problems like predictive failure analysis to leverage natural-language interaction. Personal tools he uses include NVIDIA's PDF-to-Podcast blueprint, Perplexity Pro, and Cursor.
Original sources
Podcast NVIDIA AI Podcast
Episode:NVIDIA’s Jacob Liberman on the Power of Agentic AI in the Enterprise - Ep. 250 ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.