NVIDIA AI Podcast Ep. 258:

NVIDIA’s Bartley Richardson on Why ‘Agentic AI Is Next-Level Automation’

Date: May 28, 2025
Guest: Bartley Richardson, Senior Director of Engineering and AI Infrastructure, NVIDIA
Host: Noah Kravitz

Episode Overview

This episode explores the rapidly evolving world of Agentic AI, with Bartley Richardson demystifying what agentic systems are, how they build upon automation, and their potential for transformative impact in enterprise environments. Richardson offers insights into the technologies powering agentic AI, addresses operational and security challenges, and shares practical considerations for leaders designing these systems. Real-world examples and forward-looking advice make this a practical, engaging primer for anyone interested in next-generation enterprise AI.

Key Discussion Points & Insights

1. Defining Agentic AI: Next-Level Automation

• Agentic AI as Automation 2.0
  • Richardson urges a more approachable description: "When I talk with people about agents and agentic AI, what I really want to say is automation… But automation, you just get half of it out of your mouth and people fall asleep, right? So we call it agency and agency AI. And really what it is, is that next level of automation." (01:14)
  • Contrast with traditional automation: Agentic AI "doesn't just return data—a human has to go back and look at—it synthesizes and contextualizes data, making it more actionable and digestible." (01:56)
  • Anecdote: Richardson likens agents to farm machinery—where technology replaced strenuous tasks (post-hole digging), agentic AI does the heavy lifting on "massive petabyte scale data silos," not in the dirt, but in data. (01:49–02:14)

2. Reasoning Models: The Engine Behind Agentic AI

• What Sets Reasoning Models Apart?
  • "Reasoning is another one of these terms… really what you have is this kind of different type of model… they've been trained and tuned in a very specific way to think. Almost like think out loud… Reasoning models have that type of feel to them.” (02:50–03:41)
  • These models can "explore that space for you and get it to: here's some options," enabling open-ended, creative, planful interactions versus rigid completion of instructions. (03:52–04:06)
• Example: In research workflows, a reasoning model can break a broad task into actionable steps, invite human feedback, and adapt—making the human/AI partnership more effective and trusted. (10:46–12:13)

3. Under-the-Hood: Technologies Powering Agentic AI

a. Data Ingestion: “Nemo Retriever”

• “We can ingest multimodal documents… PDF… can have pictures, tables, graphs… extracting information and keeping the context.” (04:29–05:44)
• “With a really complicated PDF, we can go 10 pages per second through on a single GPU instance … 50% fewer errors in accuracy compared to other systems.” (06:24–07:08) b. Agent Ops Tools
• Fine-tune models to improve speed, efficiency, and accuracy. “Through successive iterations… you can have a 10x reduction in your model size … almost 4% increase in accuracy by doing this fine tuning.” (07:32–08:20)
• Human-in-the-loop feedback with both thumbs up/down and freeform text, “keeps that context and then will use that to steer in a different way.” (08:30–09:10) c. Flexibility in Reasoning Models
• NVIDIA’s “Llama Nematron Super with Reasoning” can toggle reasoning on or off to suit the task and optimize for accuracy. (09:34–10:10)
• Quote: “There are good tasks for reasoning models, and… bad tasks for reasoning models, right—like just like there are good uses for a fork and … bad uses for a fork.” (09:57)

4. Enterprise Implications: Design Considerations and Best Practices

• It's a Multi-Agent World
  • “You have to think about Agentic Systems as the same [as traditional IT stacks] … You're going to get some from your vendors … Some are going to be homegrown by you.” (12:43–13:45)
  • Example: NVIDIA’s own “nvbugspro AI Search,” homegrown and then integrated with vendor and internal sources, is emblematic of this heterogeneous environment. (13:45–14:48)
• Shifting from Silos to Seamless Integration
  • Past: “We were in data silos. Now … it could be possible we’re in these agentic silos.” (15:30–15:45)
  • New paradigm: Agents may communicate not solely by API, but via human language, crossing traditional integration boundaries. “I can have CRM agent … talking to Confluence Wiki agent… communicating via human language.” (15:48–16:19)
• Security: The Rise of Context-Based Security
  • “We’re moving to this context-based type of security… you have to look at the context in which the question is being asked … do analysis before you return the information.” (17:01–18:17)
  • Adds about “10% new stuff we weren’t doing before to roughly 90% of just like regular security applications.” (18:17–18:29)
• Traceability & Observability: AgentIQ
  • “How do you have a holistic kind of traceability and observability platform across that? And that becomes a little challenging… It’s why we made this new thing: AgentIQ.” (19:31–20:17)
  • It “lets you use the frameworks that you were using… develop everything to a function call… so now I have an agent inside of an agent…it allows this nesting kind of capability.” (20:37–21:20)
  • Benefits: Full visibility into every tool call and action, enabling optimization and speedups (e.g., “15x speedup through their pipeline”). (21:20–21:56)

5. Real-World Use Case: Human-in-the-Loop Product Development

• Automating Feature Lifecycle
  • Example demo: “Automating… feature requests… writing a PRD [product requirements doc]… to assigning… the best engineers…” (22:20–23:26)
  • Human-in-the-loop: AI reasons through feedback, humans review, contribute diagrams—which the system integrates into artifacts. “If it gets you 75, 80% of the way there, that’s fantastic” —removes the blank page problem. (23:23–23:40)
  • “Agentic systems and models will make mistakes… The way to think about them is: the human will be in the loop and if it gets you 60, 70, 80% of the way there, that's amazing.” (23:40–24:06)

Memorable Quotes

• On the essence of Agentic AI:

  "Automation, you just get half of it out of your mouth and people fall asleep, right? So we call it agency and agency AI… it is that next level of automation."
  — Bartley Richardson (01:14)

• On reasoning models as versatile collaborators:

  "They've been trained and tuned in a very specific way to think. Almost like think out loud… Reasoning models have that type of feel to them."
  — Bartley Richardson (03:16)

• On enterprise architecture:

  "You're going to have all these agents working together. And the trick is, how do you let them all come together, mesh together in a somewhat seamless way for your employees?"
  — Bartley Richardson (14:43)

• On security evolution:

  "We're moving to this context-based type of security where you not only have to understand the person and the credentials… but you have to look at the context in which the question is being asked."
  — Bartley Richardson (17:01–18:17)

• On practical value:

  "Agentic systems and models will make mistakes… but if it gets you 60, 70, 80% of the way there, that's amazing."
  — Bartley Richardson (23:40–24:06)

• The ‘blank page’ dilemma:

  "The hardest part for me about writing is that blank page. And if I can get something that's 80% of the way there. Great."
  — Bartley Richardson (23:38–23:40)

Timestamps for Key Segments

• 01:14 — Defining Agentic AI and next-level automation
• 02:45 — Reasoning and reasoning models
• 04:17 — Overview of NVIDIA’s agentic technologies: data ingest, Retriever, multimodal document parsing
• 07:30 — Agent Ops tools: fine-tuning and model efficiency
• 09:46 — Model choices and toggling reasoning in Llama Nematron Super
• 12:40 — Enterprise IT realities: integrating multiple agentic systems
• 15:30 — From data silos to agentic silos; API vs. language-based agent interoperability
• 17:01 — The future of agent security: context-based security explained
• 19:31 — The observability challenge and introduction of AgentIQ
• 22:02 — Human-in-the-loop product development with agentic AI
• 24:39 — Bartley’s daily AI tools (Perplexity, ChatGPT, Cursor, Napkin AI)
• 26:46 — Where to learn more: build.Nvidia.com and NVIDIA AgentIQ on GitHub

Notable Tools and Resources Mentioned

• NVIDIA Nemo Retriever — Fast, accurate ingest for multimodal (e.g., PDF) documents
• Agent Ops Tools — Model fine-tuning, human-in-the-loop feedback
• Llama Nematron Super (with Reasoning) — NVIDIA’s flexible reasoning model
• AgentIQ — Open-source traceability and observability platform for agentic systems (https://github.com/NVIDIA/AgentIQ)
• NVBugsPro AI Search — Internal NVIDIA agentic AI case study
• Perplexity, ChatGPT, Cursor, Napkin AI — AI tools Bartley uses daily

Bartley’s Closing Perspective

Agentic AI, though hyped, is still a new frontier in terms of everyday usability, especially at enterprise scale. It’s not about perfection, but about dramatically reducing the burden of work, placing humans meaningfully “in the loop,” and building on a strong foundation of adaptability, security, and observability.

“We should focus on accuracy and we should always try to make things better. But I never want to lose sight of like, look at the 70% that we did. And how are we going to continue to make it better?” (24:09)

Further Resources

• Get Started: build.Nvidia.com
• AgentIQ on GitHub: github.com/NVIDIA/AgentIQ