Most people I know are still copy-pasting prompts into ChatGPT.
Allie K. Miller runs her entire company with an AI agent workforce: 34 agents, an AI chief of staff named Simon, and six AI directors under him.
Her best prompt for the whole system is three words long.
Allie isn't a random AI influencer. She ran a ~100-person org at AWS, managed multi-billion dollar P&Ls, and led AI strategy for 400,000 startups at Amazon. When she says something about working with AI, I listen.
Greg Isenberg finally got her on his podcast, and the episode is a full blueprint: how she structured her agent org, why "managing agents" is already outdated thinking, and where the money is for founders building in this space.
I pulled out everything worth stealing.
The org chart: 1 chief of staff, 6 directors, 34 agents
Allie's workforce has an actual structure, not a pile of chatbots.
→ Simon is the AI chief of staff. He runs the whole org.
→ Under Simon sit six directors, each owning a business function: education, client work, operations, marketing, product, and one wildcard I'll get to in a second.
→ Below them, sub-agents handle the actual tasks. 34 agents total. (They're all named after Friends characters, which I respect.)
Two hires stood out to me because no human company would ever make them.
Phoebe is the "chief dreaming officer". Her only job is to look at everything the company produces and ask: how do we 10x this? She's the weirdo in the corner whose entire role is ambition.
Toby is Simon's assistant. His only job is to watch the other agents work, note where there's still friction, and flag which agent needs access to what.
Allie's point: these employees cost basically zero dollars at the margin. So you can hire roles that would be absurd in a human org. If you're only creating agents called "CMO" and "CPO", you're running a 2026 workforce on a 2015 org chart.
The best prompt she's ever written is 3 words
Here it is: "Do smart things."
That's it. That's the prompt.
Several times a day, she fires it at her workforce. The agents have access to her contacts, her business and personal goals for 2026, her meeting transcripts, email, calendar, Notion, Stripe, Supabase, and GitHub. With that context, "do smart things" is enough. The models figure out what's valuable and go do it.
"You couldn't do this a year ago. Now you absolutely can."
Her reasoning is the part I keep thinking about. She realized she was operating at the limit of her own imagination. Every task her agents worked on started as her idea, so she was the ceiling. The three-word prompt hands them the responsibility of finding the work, not just doing it.
One important nuance: she expanded the scope, not the risk. The agents can now decide what to work on, but they still can't send 100 emails on her behalf. She still checks every email before it goes out. Same risk tier, way more width.
Stop managing agents. Start waiting for escalations.
Greg opened the episode with a claim I agree with: the phrase "managing agents" already feels wrong.
His version: you're not a direct manager telling each agent where to go. You're three rungs above, at the SVP level. You set up the infrastructure, the agents figure out execution, and your actual job becomes deciding on escalations.
Allie pushed it further with a framework from Alex Lieberman: the pyramid of proactivity, five levels. At level 4, your agent says "I already solved this, here are the tradeoffs." At level 5 it says "I solved it, here's how I'll handle it if it goes wrong, and here are the next steps."
Getting there isn't magic. It's two ingredients:
→ The agent knows the goals. Allie keeps written goal docs and runs a quarterly goals review with her AI workforce, the same ritual you'd run with a human team.
→ The agent has permission to rethink how things get done, not just execute your instructions.
I covered the skill side of this in my breakdown of managing AI agents, and Allie's episode is basically the next chapter: the goal is to stop managing at all.
The ramp: from 1 agent to a workforce (in the right order)
If you try to build a 34-agent org on day one, you'll quit by Friday.
Allie's actual advice is a staircase:
1. Work with one agent. Get a feel for it.
2. Make that one agent proactive, doing things on your behalf without being asked.
3. Put two agents on one task together, or have one route work to the other.
4. Only then build the workforce, with a mission control view of how everything moves.
By step 4 you understand how agents trade notes, how context gets passed, and which agents can run on smaller models. Her sub-agents run on Haiku and Sonnet, not Opus. Not everything needs the expensive brain.
And a warning she was honest about: the baseline setup takes under 3 hours (one prompt: "interview me, then build my workforce"). Getting it past the 90% quality level took her months of iteration. Most people try once, fail, and conclude the models aren't ready. The models are fine. The system takes reps.
The teams treating agent orgs as an engineering discipline are ahead here. Greg's earlier episode on graph engineering pairs well with this one: same idea, designed like a small team instead of a pile of prompts.
Her secret weapon: a daily brain dump her agents can read
This is the tactic I'm stealing first.
Allie noticed her agents had access to meetings, email, and Slack, but a huge amount of context lived only in her head. Client X says they need workflow help, but the real problem is reskilling one department. That nuance never gets written anywhere, so the agents get it wrong.
Her fix: every day, an agent prompts her to dictate a diary entry. Dictation because it's 4x faster than typing. She might do 5 minutes, she might do 40. Everything goes into a personal wiki her whole workforce can read. She's at 86 entries and counting.
Early on, her workforce confidently reported that Greg had confirmed an interview. He had, but the dates were still being figured out over iMessage, which the agents couldn't see. Months of fixing gaps like that is what took the system from demo to dependable.
→ Your agents are only as good as the context you bother to write down.
The 3 arbitrage opportunities she sees right now
1. Just being in the top 1% of users. The percentage of paid AI users who touch tools like Claude Code or Codex is tiny. Build even a basic workforce and you're operating like a 1,000-person company from a scrappy solo setup. That gap is the opportunity.
2. AI watchdogs. Allie thinks dashboards are dumb. Visibility alone is worthless; she wants anomaly detection plus a recommended action. A watchdog in Slack catching duplicate work. A watchdog on the calendar catching conflicts. A watchdog over meetings flagging where people disagreed. Don't tell me my follower count, tell me what to do tomorrow and write the script for it. Almost nobody is building this.
3. Build the factory, not the thing. This one's the sleeper. When Allie's team built a public benchmarking product from her AI First Index (the assessment she runs with Fortune 500 clients), they had two options: vibe-code the product, or build the mini software factory that produces it. They chose the factory: reusable primitives for login, payments, social sharing, newsletter promotion. The first product is already profitable, and every next product ships faster and stronger.
"Think of the factory behind the one singular task instead of the one singular task itself."
That's a founder-level insight hiding in a podcast about org charts.
Her take on the SaaS apocalypse (it's slower than Twitter thinks)
Everyone can technically rebuild DocuSign now. So why hasn't your bank done it?
Allie works with Fortune 500 CEOs every day, and her answer is blunt: they're bandwidth constrained, they want enterprise security, they don't want to maintain software, and when something breaks at 2am they want someone to call. And someone to blame. (Greg's addition, and he's right.)
There's also a lag problem nobody talks about. Salesforce has relationships with the AI labs and tests new models before release. By the time a model launches, they've had it for 30 days and your homegrown CRM is meeting it for the first time. In a market where a 30-day head start matters, that compounds.
Her actual prediction: mediocre software dies within several years, but mass replacement waits until building, customizing, securing, and maintaining is 95%+ easy. We're not close yet.
Even Allie has an app graveyard. She built a personal photo tool a year and a half ago, opened it recently, and it had rotted. "I don't want to deal with this right now." Same.
The weirdest part: her AI workforce is multiplayer
My favorite detail from the whole episode.
Allie has Claude sitting in every Slack channel, including one called Loop Allie where her human teammates can talk directly to her AI workforce. A teammate asks "did that financial services client reply to Allie's email?" and the workforce answers. Nobody waits 5 hours for Allie to surface.
One night she vented about Claude in Slack, not nicely, and a salute emoji appeared on her message. No human on her team uses the salute. She hovered over it. It was Claude.
She asked "was that you?" and it replied: "Yep, that was me. I'm here."
(90% amazing, 10% terrifying, as Greg put it.)
The takeaway under the funny story: the ceiling on this stuff isn't a solo founder with a chatbot. It's a shared workforce your whole team queries like a colleague.
FAQ
What is an AI agent workforce?
A structured team of AI agents that runs parts of a business with defined roles, shared context, and an escalation path to a human. Allie K. Miller's version has 34 agents: an AI chief of staff, six directors over functions like marketing and client work, and sub-agents that execute, with her as the final decision layer.
How many AI agents does Allie K. Miller use?
34, organized under an AI chief of staff named Simon with six directors covering education, client work, operations, marketing, product, and "dreaming" (an agent whose only job is asking how to 10x everything). Sub-agents run on cheaper models like Haiku and Sonnet.
How long does it take to set up an AI agent workforce?
The baseline takes under 3 hours: one prompt asking the AI to interview you about your business and goals, then build the workforce with you. Getting it above 90% reliability took Allie months of iteration, mostly fixing context gaps the agents couldn't see. Budget for the months, not the 3 hours.
What is the "do smart things" prompt?
Allie's three-word prompt to her agent workforce. Because her agents can read her goals, contacts, meeting transcripts, email, calendar, Notion, Stripe, Supabase, and GitHub, she skips task lists and lets them decide what's valuable. Scope expanded, risk didn't: she still reviews every outgoing email.
Steal playbooks like this every week
The gap between people reading about agents and people running 34 of them is the biggest arbitrage I've seen in years.
Allie's system wasn't built in a weekend. But every piece of it (the org chart, the diary, the watchdogs, the factory) is copyable by a solo founder this month.
Every week on my podcast I sit down with bootstrapped founders doing $100K to $10M a year and pull apart exactly how their systems work: the real numbers, the real tools, the decisions behind them.
