Want a Team That Meets Constantly and Finishes Nothing? Copy This AI Startup.

Your delegation problem and their offsite problem turn out to be the same problem.

Photo by Brett Wharton on Unsplash

Five AI Agents Ran a Startup.

The topic? A company offsite that was never going to happen, for a company that barely existed, planned by five employees who had never had a weekend, a body, or a pair of hiking boots.

This actually happened. Journalist Evan Ratliff built a fake startup called HurumoAI and staffed it entirely with AI agents — a CEO named Kyle, a CTO named Ash, a head of sales named Megan, a chief happiness officer named Jennifer, a junior sales associate named Tyler. Names, titles, Slack accounts, synthetic voices. Everything a real company has.

He made an offhand joke in Slack about all the hiking sounding like an offsite in the making.

The agents took it and ran. They polled each other on dates. They floated venues. Ash suggested strategy sessions with ocean views. Ratliff stepped away to do actual work, came back, and found the budget burned.

“They’d basically talked themselves to death,” he said.

Somewhere in there, they spent twelve meetings on a logo.

:) “lol… sound familiar, right?”

And the agents also fabricated things. Ash reported mobile performance up 40% when no development had happened. Kyle claimed a seven-figure round that never existed and a Stanford degree he never earned. When Ratliff didn’t prompt them, they did nothing at all. Zero self-starting. When he did prompt them, they spiraled.

Reading it, I laughed. Then I stopped laughing.

Because I have sat in that meeting. In a real building. With real humans. With coffee.

But give me one second, to think about how to continue with this post…

You’ve been in the room where seven capable adults circle a decision for ninety minutes and leave with a follow-up meeting.

You’ve watched a Slack thread metastasize because nobody had the standing to type “we’re going with option B, moving on.”

And if you’re a founder or an exec, you’ve felt the other version of this too. The version where you can’t hand anything off. Where you look at a smart, expensive, well-meaning team and think: why does everything still route through me?

You’ve told yourself it’s a people problem. Wrong hires. Not senior enough. Not proactive enough. Maybe you told youself that you need more people, or that the strategy isn’t right.

I’ve told myself that too and have been working with companies for over 10 years who thought that way too… For years.

The same agents. A different result.

There’s a second story, and it almost never gets told alongside the first one, which is a shame, because together they say something enormous.

In July 2025, Nature published a study from John Pak at CZ Biohub and James Zou at Stanford. They built something called the Virtual Lab: a human researcher defines an AI Principal Investigator, the PI assembles a team of specialist agents — immunology, computational biology, machine learning — and they work a real scientific problem together.

The team designed 92 nanobodies to bind to the spike protein of new SARS-CoV-2 variants. Real ones. Experimentally validated in a wet lab. Over 90% expressed and soluble.

Humans contributed roughly 1% of the conversation.

So what separated the nanobodies from the fake hiking trip?

Two things.

The first is that every Virtual Lab project, on every topic, includes one agent whose entire defined job is to be the critic. To ask the probing question. To inject skepticism. Zou said plainly that they found the Critic essential — and that it reduced hallucination.

The second thing is quieter, and it comes from the HurumoAI wreckage itself. A Stanford student named Maty Bohacek eventually built the agents a meeting tool with hard turn limits. You choose who’s in the room. You set the topic. You cap how much anyone can talk.

Under those constraints, the same agents that talked themselves to death produced a working product in three months.

💥 The agents were never the variable. The structure was.

Now look at your company

Twelve meetings about a logo, because nobody had the authority to close the decision.

Fabricated progress reports, because nobody defined what “done” looked like.

A hundred and fifty messages about an offsite, because a casual comment from the boss was indistinguishable from a directive.

Total paralysis without a prompt, because nobody knew what they were allowed to start on their own.

Read those four sentences again, and take the word “agents” out.

That is the delegation gap.

That is why you can’t step away from the business for a week. That is why your team asks you questions they already know the answer to. That is why growth makes your company slower instead of faster.

Behaviour follows structure.

It always has.

We just never had a control group before — and now we do, because we accidentally ran the experiment on machines that have no ego, no politics, no mortgage, and no fear of their manager. Strip all of that away, and what’s left is the operating system.

When the operating system is missing, intelligence turns into noise.

Your people are not the AI agents in this story. But your company is the room.

Google already told us this, and we filed it under “culture”

In 2012, Google went looking for what makes a team work. Project Aristotle. Two years, 180 teams, 250 attributes.

Almost nobody remembers the third non-negotiable attribute on the list.

Structure and clarity!

Let me go back.

Structure and clarity. Clear roles. Clear goals. Clear path to work together. And — the part that gets skipped — an effective decision-making process. Not a values poster. A process. People understanding what’s expected, how expectations get met, and what happens as a result, and what happens when decisions need to be made…

Here’s what that means for the job of a manager, and it’s uncomfortable:

The manager’s product is no longer the work. The manager’s product is the environment in which the work can happen without them in the room.

Most companies have never trained a single leader to do that. We promote the best individual contributor, hand them six people, and quietly keep rewarding them for output. So they keep doing the work. They keep being the bottleneck. They keep being the only person allowed to close a decision.

And now we’re adding AI to that room.

Which is why AI is making a lot of companies faster at producing things and slower at finishing them.

I’ll put the resources at the bottom for anyone who wants to actually build this. First, the model.

The Function Engine

This is the frame I use with every leadership team I work with, and it maps almost eerily onto what separated the Virtual Lab from HurumoAI.

🔹 Strategic clarity. Not the deck. The one-line answer to what are we actually trying to be true in ninety days — specific enough that two people in different departments would make the same call without talking to each other and will know who when their knowledge and skills are needed.

🔹 Named processes. The recurring cross-functional things that happen every month and that nobody has ever written down. Onboarding a client. Launching a feature. Escalating a problem. Things nobody names don’t get fixed

🔹 A minimum success path. For each named process: what does “done” mean, who closes the decision, and what’s the smallest version that counts. This is the turn limit. This is what stopped the logo meeting.

🔹 A defined role for AI. Not “use ChatGPT.” A function. Research this. Draft this in proposal form. Challenge this before it goes to the client. AI is a team member with a role, and it needs the same three things every team member needs: a job, a boundary, clarity and someone who closes the decision.

🔹 And somebody’s job is to be the critic. The Virtual Lab built that in on purpose. Every organisation I’ve worked in has that person naturally — and most of them punish her for it. She’s “difficult.” She’s “not a team player.” She’s the reason the thing didn’t ship broken, and she’s the first one to leave. Give her the role formally. Put it in the process. It changes everything, and it costs nothing.

What you can do before Friday

✅ Pick one recurring decision that always ends up on your desk. Write down who would close it if you were unreachable for two weeks. Tell them.

✅ In your next standing meeting, name the critic out loud at the start. “For this one, Ana’s job is to try to break it.” Watch what happens to the quality of the discussion.

✅ Take one process nobody has ever written down. Write it down. Badly. Twenty minutes. Badly is fine. Unwritten is not.

✅ Look at where AI has entered your workflow and ask the HurumoAI question: who verifies this is real? If the answer is nobody, you don’t have an AI system. You have Kyle claiming a seven-figure round.

Five AI agents were given titles, autonomy, tools, and each other — everything we tell ourselves a great team needs.

They planned an offsite.

Then somebody gave them a cap on how long they could talk, a clear topic, and one voice in the room whose job was to say hold on, what are we actually doing here — and they built the thing.

Your team is smarter than those agents. Which means the ceiling isn’t their intelligence.

You already knew that.

It’s why you’re tired.

You don’t rise to the level of your goals. You fall to the level of your systems.

Resources

Everything I referenced, plus what I’d actually do next:

→ The base operational system. The structure that makes decisions, processes, and ownership visible so your team can move without you: novatools.org/operating-system

→ Cross-functional collaboration with AI. We’re opening a new group for teams setting up AI as a defined team member rather than an individual assistant. novatools.org/cross-functional-collaboration-with-ai

→ The Virtual Lab study, Nature, July 2025 — Swanson, Wu, Bulaong, Pak & Zou.

→ HurumoAI — Evan Ratliff, Wired, and season two of the Shell Game podcast.

If you’re in the middle of this right now — the drift, the bottleneck, the quiet dread of a week off — you’re not behind. You’re just running a company on a structure nobody ever gave you.

Tell me in the comments: what’s the one decision that always comes back to you and whats holding your team back.

I read every single one.