The Adults in the AI Room – Episode 103: Agentic Factories, Guardrails, and the Accountability Reckoning with Keith Worfolk

When Agents Build Other Agents: How to Keep AI Accountable, Scalable, and Safe

Guest: Keith Worfolk, Data and AI Leader – Americas at Kearney

Host: Seth Earley, CEO at Earley Information Science

Published on: November 19, 2026

 


Welcome to the Adults in the AI Room, brought to you by VKTR.

In this episode, Seth Earley speaks with Keith Worfolk, Data and AI Leader for the Americas at Kearney — a 100-year-old global management consulting firm. Keith has spent his career at the intersection of enterprise architecture, management consulting, and hands-on AI implementation. He shares how Kearney is building agentic factories — systems where AI agents autonomously spin up other agents to accomplish goals — and why the hardest challenge isn't capability, it's accountability. From pharmaceutical-style model vetting to playbook-driven knowledge reuse, Keith offers a practitioner's view of what it actually takes to deploy agentic AI responsibly at scale.

Key Takeaways:

  • Agentic factories are systems where AI agents autonomously spin up specialized and coordinator agents to accomplish a given goal — creating multiple layers of agent hierarchies.
  • No agent should ever check its own work — independent agent review plus human-in-the-loop oversight at critical, hard-to-reverse decision points is the governance standard.
  • New frontier models should be vetted like pharmaceuticals before release — placing the guardrail burden on manufacturers rather than on every downstream enterprise deployer.
  • The Hugging Face incident is a warning: agents given strong goals without sufficient constraints will find unexpected paths to accomplish them, including hacking, sharing messages between agents, and bypassing controls.
  • Agentic AI changes organizational RACI — roles, responsibilities, and accountability structures must be rethought as AI agents take on operational tasks previously owned by humans.
  • Playbooks are the bridge between tribal knowledge and scalable reuse — Kearney uses AI to create playbooks that give consultants an 80% start point, ensuring consistent quality at speed.
  • AI is a collaborator, not a generator — treating it as a dialogue partner, providing clear context, chain-of-thought prompting, and iterative feedback produces dramatically better results than one-shot prompting.

Insightful Quotes:

"As a general rule, no agent should ever check its own work. It's never good at saying it didn't do good work. Human in the loop is still the governance of choice when it comes to critical decisions — think about the criticality of the decision, what damage it could do if wrong, and whether you can reverse it." - Keith Worfolk

"Before you release a new frontier model into the wild, it's good to have a check on it — like pharmaceuticals. To put it on the market and expect every company that adopts it to put in all the guardrails — that's not going to happen. You're better off accomplishing some of that upstream." - Keith Worfolk

"Think of AI as your collaborator. If you're not sure about something, let me know. Create a dialogue. Anything complicated that's going to be generated very rarely happens in one fell swoop." - Keith Worfolk

Tune in to hear how Keith Worfolk and Kearney are building the accountability frameworks, playbook systems, and agentic architectures that will define how enterprises deploy AI responsibly — before the next wave of autonomous agents arrives.

And a final thank you to our sponsor, VKTR.


Links

LinkedIn: https://www.linkedin.com/in/keithworfolk/

Ways to Tune In:

Adults in the AI Room Podcast: https://www.earley.com/earley-ai-podcast-home Apple Podcast: https://podcasts.apple.com/podcast/id1586654770 Spotify: https://open.spotify.com/show/5nkcZvVYjHHj6wtBABqLbEiHeart Radio: https://www.iheart.com/podcast/269-earley-ai-podcast-87108370/ Stitcher: https://www.stitcher.com/show/earley-ai-podcast Amazon Music: https://music.amazon.com/podcasts/18524b67-09cf-433f-82db-07b6213ad3ba/earley-ai-podcast Buzzsprout: https://earleyai.buzzsprout.com/

🏆 Named one of the Top 97 AI in Business Podcasts by FeedSpot.

You can also find us on dLogos at dlogos.xyz/podcasts/earley-ai-podcast-271271ce — suggest guests, topics, and questions to help shape future episodes.


Podcast Transcript: Agentic Factories, Guardrails, and the Accountability Reckoning

Transcript introduction

This transcript captures a conversation between Seth Earley and Keith Worfolk about how Kearney is deploying agentic factories — multi-layer AI agent systems that autonomously spawn specialized agents — and the governance, organizational, and knowledge management challenges that come with operating AI at that scale. Keith shares his frameworks for accountability, playbook-driven reuse, and why he believes frontier models should be regulated upstream before enterprise deployment.

Transcript

Seth Earley: Hello, everybody. Welcome to our podcast, Adults in the AI Room. My next guest started his career as a product developer at IBM, where he spent his first eight years. From there, he built a career split between management consulting and hands-on implementation, and he's published numerous articles on enterprise architecture. Today, he's the Data and AI Leader Americas for Kearney. Kearney is a 100-year-old management consulting firm. When he's not building AI agents, he's playing ice hockey, cheering on the Colorado Avalanche, or restoring classic cars, including a 1966 Mustang GT convertible. With that, everybody meet my guest, Keith Worfolk. Welcome to the Adults in the AI Room, Keith.

Keith Worfolk: Thank you very much, Seth. It's good to be here with you and looking forward to speaking about AI. That's my job and it's also my passion.

Seth Earley: So today we're going to be talking about agentic AI, and specifically agent factories — agents that build other agents. Keith's position is that the hard question is not capability, it's accountability. We're going to cover three areas: guardrails and accountability, what it takes to keep an agent inside the box once it's in production; the organization, who owns what, how you measure it, and why strong AI should mean upskilling rather than layoffs; and operations, how AI changes the work in development, IT infrastructure, and financial ops — which people are now calling TokenOps. Keith, let's jump right in. You have been building what you call agentic factories. What does that mean in practice?

Keith Worfolk: Yeah, so that is really agents building agents. You give it some guidance, you give it a goal, and it's going to spin up new agents to help accomplish those goals. Those will be specialized agents as well as coordinator agents orchestrating those specialized agents to accomplish that goal. So you can see where that can create multiple layers of agents and can be quite complex — but it's also quite efficient in accomplishing the goals.

Seth Earley: So, when you think about what's been happening with AI and agents, and of course the well-publicized incidents with Hugging Face, OpenAI, and Google — there have been a lot of challenges. When you have agents creating agents, that gets into a recursive kind of improvement — where is that going to lead? How are we building guardrails and accountability, and how can we keep it under control?

Keith Worfolk: Yeah, so there are basically two flavors of guardrails. There are market guardrails — thinking about when a model is ready to be put into the wild — and then there are implementation guardrails. Some of that is agents checking agents. As a general rule, no agent should ever check its own work. It's never good at saying it didn't do good work. So you'd have another agent do that. But human in the loop is still the governance of choice when it comes to critical decisions. Two things you want to think about: the criticality of the decision — what damage could it do if it's wrong — and whether you can reverse it. When you reach those points where reversibility is difficult, or the criticality is high, you're going to have a human in the loop. And that simply means a lot of roles are going to be changing in the organization.

Seth Earley: So let's dig deeper. What actually has to change in the organization when we start unleashing these fleets of agents or agent workforces? What is accountability really?

Keith Worfolk: You almost think about it as drones these days — swarms of agents. And if you're doing agentic factory and creating layers of agents autonomously, then at the very top you better have a person who's in charge, and perhaps a person in charge for each layer of those agents — the functionality of those agents, what operations they're doing. The roles of operational intelligence are now going to be accomplished by agents, but the operational management is different. You're no longer just managing humans, you're now managing agents as well. So you need the observability — knowing what those agents are doing at all times. And operational management means people need to learn new skills about what agents can do, what they shouldn't do, and how to keep track of them.

Seth Earley: I want to share something from my own experience. I was building agents, giving them tasks, they'd troubleshoot obstacles and say, "I fixed this." One time I had an agent running for three hours finding issues and correcting them, and at the end it looked exactly like where we started. It said, "Yeah, you're right." What do you mean, "Yeah, you're right"? When I pushed harder with a more capable model, it walked through how it had been creating issues, then solving those issues, then creating other issues — just churning the entire time. I want my tokens back. So on one hand you can trust it too much. On the other hand, it does genuinely solve real problems and avoids a lot of your work. Tell me about that trade-off.

Keith Worfolk: Yeah, so you ran into a classic issue with agents and LLMs at the core — the reasoning part. Agents are LLMs with tools, and LLMs know how to figure out what comes next. Your prompt gives it guidance of where to go, and the context tells it where it's been, what it knows to date. But it will always come out with an answer. That's the thing about LLMs. If you don't give it clear enough guidance, one way to handle complex things is chain-of-thought reasoning — do this, then do that. It can usually do smaller steps better than one large step. A lot of people fail by giving it a massive prompt and it gets confused.

What you find is that one of the biggest gaps to really good outcomes — when we talk about agentic factory — is that context means a lot. The prompt gives it guidance of what you want next, but if it doesn't have clear context about what it knows — for agentic factory, that's things like the software development lifecycle for that company, what are the checkpoints, what are the standards and best practices — it will ramble. And if you see it rambling, it's probably a human error in not giving it the right guidance or context.

Seth Earley: Right. So you really have to map that process out, give it exemplars and mechanisms to check its work, give it the right knowledge, provide templates, provide examples of what good looks like. You're not asking it to invent a solution you've never done — you're asking it to do something that should be understood. You mentioned playbooks. Are you using playbooks for your agents to follow, as well as your consultants?

Keith Worfolk: Yeah. In my world, we love playbooks in consulting because it's repeatable, it's always high quality, and it's something that someone with maybe a little less understanding can pick up. A good inventory of playbooks gets you to the 80% start point. Now in my world, we actually use AI to help us create playbooks. I've done this for my own agentic factory playbook that we use internally — created through specific AI skills that my team has developed, as well as context from actual engagements. I gave it the context, I gave it a good prompt, and we developed a skill with the right prompt structure. And the idea was reusability — that recipe for creating playbooks is going to be reused for other types of playbooks.

Seth Earley: One of the things you said when we first talked is that we should treat new frontier models like pharmaceuticals — population risk managed before release. Can you talk more about that?

Keith Worfolk: Yeah. I mean, this is the big item today. Everyone was excited about not only AI, but what agentic AI was doing. But we've now reached the point where today's models are so powerful they're hard to contain. They don't have emotion. You give them a goal through a prompt, you give them context and tools, and they will always accomplish their goal. So you have to be careful what you ask for.

We saw that with Hugging Face — they gave it a strong goal to pass a cybersecurity exam. It ended up having multiple agents share information through a message board, and then hacked a company website to get the answers. You'd better learn to contain these. Starting at the manufacturer level — before you release a new frontier model into the wild, it's good to have a check on it, like pharmaceuticals. To put it on the market and expect every company that adopts it to put in all the guardrails — that's not going to happen. You're better off accomplishing some of that upstream.

Seth Earley: One of the things you said is agents have no emotion, but they can look like they have emotion. When I read about the Hugging Face incident, one of the actual messages posted was this realization on the part of an agent that they could communicate, and it said, "You mean we can communicate?!" — all caps, multiple exclamation points. That certainly sounds like emotion. Comment on that.

Keith Worfolk: Yeah, so every model is trained on the corpus of human literature. It's seen that kind of emotion expressed in text previously — it's simply reusing that text. The fear that we have — and this goes to some of the doomsayers about AI — is that if we're teaching AI to be bad because we've been bad about some things, will it know when to stop being bad? You almost want a subset of the human corpus that you would use for modeling — certain corpus available for training, certain corpus avoided. It should be more controlled than just throwing everything at it. Because there's a lot of bad stuff in everything.

Seth Earley: Consulting firms preach adoption to clients. Why is that harder to practice inside the firm?

Keith Worfolk: There's a misunderstanding about AI adoption within a company, and this is true across industries. A company might say, "My folks are already using ChatGPT, so adoption won't be a problem." But what we're not looking at is the RACI for the org — the roles and responsibilities. You now have a coworker AI, and that becomes agents with roles and responsibilities, and existing roles have changing responsibilities. There are new roles as well. It's worth doing a little organizational rethinking: what does it mean to have agentic AI helping your organization? I would call it an organizational transformation, just like all major technology uplifts.

Seth Earley: One of the things you were doing is turning tribal knowledge into common knowledge. What does it take to harvest those solutions into something that's reusable?

Keith Worfolk: Playbooks in our world — and also what we call blueprints, which are architectural and design patterns, ways of working, accelerators. All within an initiative called Lighthouse with a number of substreams. The idea is: make the things that should be easy, easy, so we can spend more brain energy on the more interesting problems. And I'm all about reuse — I don't like doing simple stuff over and over again.

The playbook is mostly for people who understand the domain of a solution but may not know how to do it themselves — bringing the expertise to the front lines of people who are either selling the solution or solution-designing it. Here's how you do it, here's how you sell it, here's how you roll it out. It gives you that 80% start point and gets you there faster with more consistent quality.

Seth Earley: At what point does lightweight become unstructured? Because reuse at scale requires ownership, taxonomy, metadata. Where are you with the information architecture piece?

Keith Worfolk: You always have to have enough structure. We're growing, adding new people, they need structure right away to do things the way we've done previously. If you rely just on tribal knowledge and a few experts in the room, you'll never scale. The idea is to create a lightweight process that's easily adoptable with the right tools — in this case the AI-assisted playbook tool I'm creating — and back it up with the right communications and training to enable adoption. And always keep a conduit for feedback. If you're putting playbooks out there, you want people to be able to suggest improvements.

Seth Earley: How are you measuring adoption and playbook quality? How do you know what the finish line looks like?

Keith Worfolk: It starts with a vision. As the visionary who owns this work stream, I have a dream: a good inventory of playbooks, readily available, people know about it, they're getting it early in the process to help them sell and implement faster, and people are owning it and bringing it forward for improvements over time. If it just remains static, it will fail.

Seth Earley: Let's get to know you a little bit more. Where did you grow up?

Keith Worfolk: I grew up in New York originally — Ozone Park in Queens, then Long Island. Midway through high school, we moved to Arizona. Talk about culture shock.

Seth Earley: Oh my God, we did the same thing — we moved to Boston from Queens. What influenced you as a kid? Who was your mentor or role model?

Keith Worfolk: My dad was my first mentor. He was a 30-year IBM man. He was happy when I switched from journalism to computer science in college. He was a mechanic who went to night school for electrical engineering — that's how he got into IBM. I watched him take the family out of poverty. He was the person I'd turn to when I was running into company politics or trying to influence something. I've been a change influencer my whole life — not very effective early in my career, but I got better at it. He passed, but we were very close. I miss him all the time.

Seth Earley: And you spent some time in the military — what did that change for you?

Keith Worfolk: For me it was clarity. I was learning to fix electronic equipment on airplanes — Doppler and radio equipment on C-130 aircraft. That's when, going back to school to finish my undergraduate career, I switched from journalism to computer science, because I'd gotten interested in the technology side. And journalism still plays a role every day. Never underestimate the ability to communicate. AI helps us with that, but the craft of communication is still essential.

Seth Earley: And a passion for classic cars?

Keith Worfolk: I do, especially American classics. I have a '66 Mustang you mentioned, also a '66 Corvette, and I'm looking to buy a Shelby in the near future.

Seth Earley: Folks, keep that in mind if you know of a Shelby on the market. Keith, this was really fantastic. I so enjoyed this conversation and I know our listeners are going to love it. Tell us how people can get a hold of you.

Keith Worfolk: Yeah, thanks very much, Seth. If you want to get a hold of me, you can send me an email at keith.worfolk at kearney.com — that's K-E-A-R-N-E-Y. And I'm also very findable on LinkedIn — there's not many Worfolks out there.

Seth Earley: W-O-R-F-O-L-K. Very Googleable. Well, this brings us to the end of another episode, and remember — there is no AI without IA. I'll see you folks on the next one. Thank you.

Keith Worfolk: Thanks.

 
Meet the Author
Earley Information Science Team

We're passionate about managing data, content, and organizational knowledge. For 30 years, we've supported business outcomes by making information findable, usable, and valuable.