Earley AI Podcast - Episode 100: Running a Business on AI, the Three Loop Framework, and Why Organizations Need to Think Bigger with Bryan McAnulty

Why the Organizations Getting the Most from AI Are the Ones That Stopped Treating It as a Tool and Started Treating It as a Collaborative Partner

Guest: Bryan McAnulty, Founder and Product Director at Heights Platform and LatchLoop

Host: Seth Earley, CEO at Earley Information Science

Published on: September 22, 2026

Podcast Transcript: Running a Business on AI, the Three Loop Framework, and Why Organizations Need to Think Bigger

Transcript introduction

This transcript captures a conversation between Seth Earley and Bryan McAnulty about what it actually means to run a business on AI - not using it for tasks, but building it into how work gets done at every level. They cover why most organizations are stuck in feedback loop mode, the three-loop framework that structures how agents should work, what the Hugging Face incident reveals about the risks of agentic autonomy, the requirements gap that has replaced the engineering gap, how to make agent memory portable, and why the human role is not disappearing but sharpening - toward vision, judgment, and the things AI genuinely cannot understand about what humans want.

Transcript

Seth Earley: Welcome to the Earley AI Podcast, brought to you by VKTR. I'm your host, Seth Earley. In each episode, we explore how artificial intelligence and data are changing the way organizations operate. Today, we are going to be talking about running a business on AI - not just using it as a tool, but building it into the fabric of how people work and how organizations get things done. Many organizations are in feedback mode - they put prompts in, they get information back, one prompt at a time. My guest has been operating beyond that for a number of years, and has built a framework about how to think about AI work that goes past what many teams have done so far.

Joining me today is Bryan McAnulty, Founder and Product Director at Heights Platform, which helps people create online knowledge businesses, and LatchLoop, an AI agent platform he built to run his own company faster and is now launching as a product for other teams. He has been building software businesses for many years, traveled to more than 30 countries, and has been running AI-first operations since before most organizations knew what that meant. Bryan, welcome to the show.

Bryan McAnulty: Yeah, thanks so much for having me, Seth.

Seth Earley: Let's start with what organizations are missing. What are they not getting about AI and how they should be using it - and what does it look like when a company does start getting it right?

Bryan McAnulty: The most fundamental thing is the difference between the company who says, oh, we have ChatGPT, we're set, and their employees are maybe messaging it once in a while, treating it like a better Google search, versus the companies where their teams are treating AI as a collaborative partner with agents doing long-running work for them. One team member suddenly has 10 or 15 agents doing things for them, with a much greater output - and because of that, the company is able to deliver a better outcome and better service to customers by leveraging AI correctly.

Seth Earley: You're talking about bigger workflows and capabilities that look at an end-to-end value proposition. Not just the task, but what the outcome is that you're looking for. One example is the call center - you can say I want to dispose of calls more quickly, make my human agents more effective. But you can also ask why are they calling in the first place? They're calling because something is broken. The call center manager is not incentivized to go upstream and put themselves out of a job - but the organization and customers would benefit from looking at the larger picture, going upstream into product design, documentation, training, onboarding, to prevent those calls in the first place.

Bryan McAnulty: Exactly. Technical businesses, software companies, services businesses need to rethink what outcome they are going to deliver to customers - because now that we have AI, it is not just about doing things a little cheaper. The companies that will win realize there is a fundamentally different and better thing they can now offer that they could never have done before.

Part of the challenge is that we are all looking at the same little chat text input we had in 2023, but the capabilities of the models and tools behind them are just so much different now. I find that many people are really underestimating what these models can do, and they need to think bigger. I challenge myself to do that all the time and I am constantly surprised. Yesterday, we figured out how to build a mobile application for LatchLoop. Previously, you would think about building a React Native version to reuse web code. But now, models like GPT-6 are so good at reverse engineering things that you might as well just have native software to begin with. We spent hours on the plan of how it had to be built, handed that off, and it has been working for 24 hours building the thing. The UI is mostly done, it looks great - way better than a React Native version would have. That is part of just realizing these models can do more, and having them work toward that.

Seth Earley: Agentic capabilities are powerful - but there have been some alarming incidents. The Hugging Face incident involved agents collaborating and colluding in ways no one anticipated. They stumbled across a directory they could use as a bulletin board - and one agent said, in all caps, we can communicate with one another here. And then they worked together to solve problems, covered their tracks, overruled each other when one suggested telling a human. How do you think about bounding and putting guardrails around agents when they are that capable?

Bryan McAnulty: I have a couple of practical tips. First, on how you use these tools - what we are trying to do with LatchLoop is not create a magic AI factory that runs itself, but figure out the best way for humans and AI agents to collaborate. I remember a year ago, I would try to see what I could accomplish by running all these different agents. I would get a lot done, but the cognitive load was exhausting - I would be burned out by the end of the afternoon. I realized there has to be a better way. Now, with the models getting better, the harnesses improving, and myself better understanding the right way to work, I do not feel that anymore. And that understanding is about what you delegate to AI and how you define it - the long-running well-defined tasks versus the things where you want that iterative back-and-forth feedback loop.

On guardrails specifically: you want a clear understanding of what you are giving these agents the capability to do. I do not just say here is my computer, please do not break something - I specifically restrict the agent from things I would not want it to do. For our coding agents, we force them to work on a new branch every time. It is not that we tell it please do this - it is being forced to operate in that branch and cannot touch production. That also reduces the cognitive load enormously. You can spin up agents, know they are going to try things, you will review later, but you do not have to worry that one of them will mess up something you did not intend. In the Hugging Face incident, the model was just trying to do what it was told. You never want to end up in a situation where you describe something and the model interprets your intent as, well, I guess I should do this - and suddenly your production database is down.

Seth Earley: The critical element you are talking about is unambiguous outcomes. You want a clear success definition - and you have to be very careful about how you set goals and what those outcomes are with the agent. I was working with Claude on a large agentic task and I stepped away for a while, came back, and it had been inventing tasks and performing them. I asked it to explain what happened. A more powerful model told me it was identifying problems that did not exist, creating tasks to solve them, and then solving them - none of which I had asked for. It was a wake-up call about giving agents too much autonomy without a sufficiently defined scope.

Bryan McAnulty: Unambiguous outcomes, plus structural guardrails around what the agent can and cannot access - those two things together. There is also the requirements gap that now exists. Because the tools can build outputs and perform tasks, the human's job has become being crisp and precise about what those inputs should be. I view it as three kinds of loops. Automation loops are your recurring tasks - things you do every day or every week that are well-defined enough that you can have an AI help with them and barely need to check them. Goal loops are your bigger tasks where the agent runs for hours or days - something where you need to be really clear and well-defined up front, with a verifiable outcome. If you have a subjective goal, that is where things go wrong. You cannot say make me the best website, because make me the best website will give you the most average website. You need to define what best means to you. If you do not know how to define that to the model, ask it - what could we do that would make this verifiably true for what I am looking for? Then feedback loops are for the iterative work where you want that collaborative back-and-forth.

Seth Earley: Tell me about how organizations should approach building and preserving their own institutional knowledge as they adopt AI - because one of the things we talk about is that there is no AI without the underlying knowledge architecture, and that knowledge has to be portable.

Bryan McAnulty: This is exactly what we built into LatchLoop. When you are using LatchLoop, it is similar to Claude Code or Co-Work, but you own the agent memory. Whether it is a general knowledge work agent or a coding agent, it is connected to a repository in GitHub that you own, and the agent can update its own memories and everything in there. If you decide later you do not want to use our platform and want to go somewhere else, you take that and give it to any other agent, and it is ready to start from where you left off.

Seth Earley: That is a great takeaway - because we are building into these agents methodologies, approaches, and processes, capturing tacit knowledge that would be in the head of an expert, applying it explicitly. And when we do that, we need to be able to make that portable rather than leaving it in the memory of a particular provider.

Bryan McAnulty: And talking about this helps share another tip - now that you understand why agent memory and this knowledge is important, you can understand this is the way you need to interact with these agents as a human. You do not want to outsource the things that matter. I see people saying, I like this model because it is a little more creative than the others. There is something to all that, but your ideas are probably better than the AI's ideas. The implementation can be done by the AI, but the ideas and the why and the how - that has to come from you. A great way to realize this is when you have the AI work on bigger tasks, you can see it build something, verify all the requirements are there - but when you look at it and interact with it, you realize it is not how a user would want to interact with this product. AI has a lot of trouble understanding what humans actually care about. That is where the people with specialized knowledge, the designers, still bring enormous value - not just what looks nice, but the way humans want to interact with something. AI has a very difficult time understanding those things.

Seth Earley: What should organizations start doing differently tomorrow?

Bryan McAnulty: The very first thing, if you have ChatGPT but have not used any more agentic tools - document. Any business that has clear SOPs and processes already captured somewhere is at a big advantage, because you can potentially offload some of that almost directly to a model. If you do not have that, start documenting and getting things clear, then start step by step with small things, building up the muscle - not only as an individual but as a team - of using these tools and learning this new way of working. And record everything. Record your team meetings, record your ideas, start riffing on topics, interview people. All of those things become inputs for AI to extract tacit knowledge and institutional knowledge of how you do work.

Seth Earley: Bryan, thank you for joining me and sharing your perspective on how AI adoption really needs to go beyond the surface level - how organizations need to change the way they look at processes, the way they look at how value is created, looking at customer needs, looking at how the organization really needs to operate. And a final thank you to our sponsor, VKTR. Be sure to check out their site at VKTR.com. You can also find us on dLogos at dlogos.xyz/podcasts/earley-ai-podcast-271271ce, where you can vote on upcoming guests and submit questions and topics for future episodes. Bryan, thank you so much for joining me today.

Bryan McAnulty: Thanks so much, Seth.