NVIDIA's Bartley Richardson on Why 'Agentic AI Is Next-Level Automation'
Recorded live at GTC 2025, NVIDIA Senior Director of Engineering Bartley Richardson argues that "agentic AI" is best understood as the next level of automation rather than a wholly new paradigm, built on years of compounding work in retrieval, fine-tuning, and reasoning. He walks through the components of NVIDIA's agentic stack: NeMo Retriever for multimodal document ingestion (with claims of ~15x throughput over leading systems and ~50% fewer errors on complex PDFs), Agent Ops tools as a fine-tuning flywheel that can shrink models by ~10x while lifting accuracy ~4%, and the Llama Nemotron reasoning model with a toggleable reasoning mode. He emphasizes that reasoning models' biggest value in enterprise settings is enabling early human interaction, which improves perceived accuracy even when raw accuracy gains are modest. On the enterprise side, Richardson frames agentic systems like IT software portfolios: organizations will mix vendor agents, homegrown agents, and embedded agents across CRMs, coding repos, and wikis, creating potential "agent silos" rather than data silos. He positions NVIDIA's newly released Agent IQ as a framework-agnostic observability layer that treats every interaction as a function call, enabling traceability across heterogeneous agent frameworks. He also discusses the security shift from perimeter- and application-based models toward context-based security, where the context of the request and user determines what data can be returned. Richardson closes with a pragmatic "70% rule": agentic systems will make mistakes and rarely be 100% accurate, but the productivity lift of automating 60–80% of work is transformative, especially with humans in the loop. He demos a use case where a reasoning model drafts brainstorming questions, a human team holds a meeting, and the system ingests whiteboard photos and transcripts to produce a near-finished PRD. He recommends build.nvidia.com and the open-source Agent IQ on GitHub as starting points.
Original sources
Podcast NVIDIA AI Podcast
Episode:NVIDIA’s Bartley Richardson on Why ‘Agentic AI Is Next-Level Automation’ - Ep. 258 ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.