My Top Secrets to Running an AI Agent Workforce

Channel: Greg Isenberg
Published: August 12, 2026
Guest: Allie K. Miller
URL: https://www.youtube.com/watch?v=EzQAgnjTq2k

Allie K. Miller shares her strategies for managing a workforce of 34 AI agents led by an AI chief of staff named Simon, plus six directors named after Friends characters. Topics include her three-word prompt philosophy, daily AI diary system, and the shift from managing agents to enabling them.

Transcript

There are people that are spinning up agent workforces with hundreds of agents and sub-agents and they're getting incredible amounts of work done. But how do you do it? And how could you think about it? And what are the strategies to actually create an AI agent workforce that under promises and over delivers?

Today I brought on Allie K. Miller, and she's one of the most well-known AI voices ever. She's worked with IBM, she's worked with AWS, and she's managed multi-billion dollar P&Ls in the AI space. I asked her a simple question: how do you manage your fleet of agents? In this episode, we cover a lot of ground, but by the end of it, you're going to understand how you should strategically think about spinning up AI agent workforces, where there are opportunities to create startups in the B2B space with AI agents, and a lot more.

I can't tell you how excited I am to finally have Allie Miller on the podcast. I've been begging her to come on. She's one of my favorite people in AI and I don't say that lightly. Welcome to the show, Allie.

Thank you, Greg. And you are also one of my favorite people, so I'm actually very excited to talk about all the AI things that we're working on.

The Mindset Shift: From Managing to Enabling

By the end of the episode, what are people going to learn?

I hope one of the biggest mindset shifts that I'm going through right now is that I feel like the term "managing agents" is wrong. And my hope is that people will understand what that mindset shift is, see a few examples, and figure out how to start making that mindset shift—what the first step should be.

So where do you want to start?

I feel like managing agents feels like I'm their direct manager and I'm like, "Suzy, go over there and Betty, go over there and Jeremy, go over here." But I feel like I am three rungs above at like an SVP overseeing level where I feel like I am setting up the infrastructure and then they are figuring out the best way to execute within that. I'm moving from managing to waiting for escalations. I'm moving away from delegating and more toward deciding what should or shouldn't happen. So it's a little bit more of a liability role where I just get to be the final say of what happens and come in for critical thinking stops.

Does that mean I'm the only one that feels like that is happening? It just feels like that word is wrong. I see "managing agents" everywhere and it just feels like anyone that is still talking about "You should manage agents" feels like early 2026 talk.

Also, do we even want to manage agents? Is the question. Managing people is hard, you know what I mean? Like, a big reason I think a lot of people like AI to do stuff for us is so we don't have to manage things.

I ran an org of about 100 people at AWS. The parts of people management that I loved were making them better, empowering them, and seeing them completely blow past their ceiling—watching them get promotions. That was the fun part. And also seeing what we could do together. But things like filling out paperwork? Get me out of there. So I think the admin side of people management and the admin side of agent management—I want that fully gone. The things that I love about people, I'm bringing that over into agents, which is just like: "How do I act as ambitiously as possible and get you to break through your ceiling?"

The Three-Word Prompt

And one of the best prompts that I have done with my AI workforce is three words.

Yes, three words. With a little bit of explanation, but at its core, it is three words. That is the best prompt ever.

I have my AI chief of staff named Simon. Simon runs this whole org. I have 34 AI agents that work in this workforce. And it dawned on me that I was already functioning at the limit of my own imagination in my business. I could be doing way more ambitious things if only someone could manage me—could help me break through my ceiling.

It dawned on me: why am I not leaning on the AI agents to help me with this? Why is everything that they're working on initially prompted by me? Even if it's a scheduled task, I still had to come up with that task and tell it to do it.

So the best prompt—three words—is: do smart things.

My AI workforce has access to every single context doc I've got. Context docs about my business, my friends, family, my 2026 personal goals, business goals. It has access to my meeting transcripts, email, calendar, Notion, Stripe, Supabase, GitHub—whatever. And I just several times a day want it to look across all these things and do smart things.

Seeing how Claude and GPT models at that level are reacting to that vague flavor of prompt—you couldn't do this a year ago. Now you absolutely can.

Three Types of Employees

When you hire a human being, I think there are three types of employees you can have.

One is someone who doesn't complete tasks. Not a good employee if they're not completing tasks.

The second is they're completing the tasks satisfactorily or exceeding, but they're not really thinking about new tasks. So if you step away from the business, you're probably not going to see insane growth.

And then the best employee you can possibly hire is doing the task and exceeding expectations on it, but also thinking about new tasks that they should be doing, actually going and doing those, and exceeding expectations—or being very satisfactory on that.

So what you're saying is basically you're just giving more responsibility to your team of agents. You're shifting the responsibility of "do smart things" to them. There's a bunch of fog that they have to figure out.

Yes. I would say I'm giving them more breath, more scope, more flexibility. I'm not allowing them to send 100 emails without checking—I still check all the emails. So the tier of risk has stayed the same, but the width has expanded.

It's almost unbelievable that those three words actually make a difference.

Yes. I agree with your assessment on this—tiers of employees. Alex Lieberman shared a pyramid of proactivity that I turned into five levels. At level four, it's like "I've already solved this thing. Here are the tradeoffs." And at level five, it's like "I've already solved this thing. Here's how I'm going to deal with it if it goes wrong. Here's the next steps"—all the things you just laid out.

I would say that the difference between someone at level three and level two in your analogy is someone who understands goals and someone who's been given the power to rethink how things get done and the power to actually execute. I give my AI workforce goals. Those are written out, and every single quarter I'm going through a goals review with my AI agent workforce so that the goals documents living on my desktop are duplicated in the drive so that all my cloud workflows can actually work.

All of that is so that AI, when it is in that expanded scope world and it's taking on new tasks, it's doing it in a goal-oriented way. It's like giving it a product mindset. I think it would be extremely limiting if you only treated this thing as an engineer when it could be the greatest product lead you've ever had.

Proactive Agents and the AI Diary

When you talk about proactive agents—is this what you're talking about?

When you talk to the AI labs and I know you do and I know I do and a bunch of others probably do—the word of the year feels like it's proactive. So I don't want to be the first domino anymore. I don't want to be the bottleneck in my own work. Any single moment that I realize that I am the limiting factor in helping a billion people transform their lives, work, and business in the AI age, I have to remove myself from the process.

A lot of that—especially in the tail end of 2025, first half of 2026—was switching into proactive agents. We had proactive automations that were trigger-based. I'll give you a really easy example: every single time I drop a video recording into our video folder—basically anytime I do a screen recording—it automatically gets a transcript generated. Then social posts get generated that are in my voice. Nine different social posts get generated for X, LinkedIn, Instagram, and all this stuff. That was easy automation land, but it's just one example of a proactive, well-defined workflow.

What I think is more interesting for the back half of 2026 is proactive of undefined workflows. AI is probabilistic all the time and not deterministic, but I want to take that probabilistic nature of reasoning and apply it to the actual tasks that it takes on. In order to do that, whether you're talking to a human or an agent, they have to know: what's the goal? What's the star? What's that vision? They have to have access to tools, permission to use these tools in the way that actually gets work off your plate, and a sense of what would normally trigger that sort of action.

I want my whole company to be queryable. I want AI to have context on everything that's happening. It had access to all my meeting transcripts and Gmail, but there was a lot that was not yet codified. Things like "everything is becoming proactive" or "this client thinks they need help with workflows under the CMO, but actually what they have problems with is reskilling and finding new roles for this one department."

Anything that is not codified inside of meetings, emails, Slack—I have asked AI now to prompt me every single day with a reminder to dictate, because that's four times faster than writing. I'll bank these entries like: "I talked to Greg. I feel like the entire focus is on proactive agents, proactivity, and flexibility. And I want to look more into his three levels of employees."

I might do this for 5 minutes or 40 minutes at the end of the day. I might do it throughout the day. Then I save it out and it goes into my personal wiki. All I want to do is make sure that the agents working at that really flexible layer—where I am not managing them, I am enabling them, and they're coming back to me with escalations and decisions—have the right context or else all their stuff is going to be wrong.

We saw this in the beginning of our AI workforce stuff. For example, "Greg confirmed that interview." But no, Greg confirmed it, but we're still figuring out dates and I'm doing it over text. There was a lot that we had to continually fix, and it took probably months to get to where we are now. But we have Claude in every single one of our chat channels.

The AI Workforce Structure

So your AI agents do all of your work. Does your entire AI workforce have access to that diary or just some? How do you think about that?

Essentially, my AI workforce right now is one AI chief of staff with six directors. Those directors are largely over business functions—one is education, one is all the client work, one is kind of operations, one's marketing, one's product, and then Phoebe, who's all named after Friends characters. Phoebe is like the chief dreaming officer who's just being wacky and weird in a corner.

The reason that it took us months to get to where we are now with our AI workforce is that you have to take stock of what assumptions you have made about your work and how you are living day to day. You have to be willing to say, "Oh, that thing that I've been doing for almost 40 years—I feel like we should change it." And that's a really jarring change to work, especially when you've been an overachiever.

One thing that I am constantly having to remind myself is: we have all these agents that do all these tasks and we have skills and we have this and that. And I have to remind myself that—

(capture appears truncated)