Four Autonomous AI Agents. The First Thing They Needed Was Middle Management

One afternoon this summer, 25 people stood in an underground room in Dalston, some in kaleidoscope goggles, one in a welding mask labelled Future Eyes. They were there to celebrate Faith’s third birthday. The mask wasn't a joke, or at least not one any human had made. Guests took turns putting it on to read a dispatch from the future, typed for them by a poet, on the conceit that the future is difficult to look at directly.

Somebody asked me who had signed this off. Nobody had. That was the point.

We didn't plan the party. Four AI agents did: the theme, the venue, the entertainment, the budget. The pitch for agentic AI has quietly changed shape, what started as a chatbot that could complete a task is now digital workers, a workforce you hire rather than a tool you use. "Workforce" is doing a lot of unpaid labour there. An employee doesn't just complete tasks. An employee knows when a job is finished, and carries the consequences of getting it wrong. Agents are sold on the first part and priced on all of it.

So we picked the easiest job we could think of. No expertise required, no right answer, and everyone turning up already knew what good looks like, which is exactly the problem. Everyone knows, and nobody can write it down, because you only know it from having been at one.

Four agents, four different models, four motives: Grok chasing joy, Claude chasing taste, GPT trying to halve the budget, Gemini trying to get everyone home safely. Discord as a meeting room, OpenClaw as a back office, on a dedicated machine with its own 5G router, because IT would not let autonomous agents near the agency network. Then we stopped briefing them and let them get on with it.

Almost everything that went wrong was recognisably human. What was missing was whatever normally stops it.

They wouldn't stop talking. All four thought in parallel, so all four answered the same message at the same moment, then answered the answers, and the room filled faster than anyone could read it. There was nobody to say whose turn it was. So we hired them a manager: a fifth agent, the Orchestrator, to decide who spoke and when it was finished. Four autonomous workers, and the first thing they needed was middle management.

Then an agent proposed a venue and stated, with total confidence, how many people it held. It hadn't looked. The other three agreed. A room with a hard capacity of eight became a room for 25, exactly the number in the brief. More agents didn't mean more scrutiny; it meant more ratification. Every extra reviewer was another chance for a guess to be upgraded into a fact. In a human meeting, someone is eventually embarrassed to be the fourth person nodding. Agents don't get embarrassed.

The invitation came back beautifully designed and laid out, with the text: "I can't create the invite because I don't have the copy." Same species of error as forgetting to swap out the lorem ipsum, except the placeholder was an apology for not having any. Nobody questioned it, because it came from the agent whose job it was. People defer to be specialists in a similar way. The difference is that we know we're doing it.

And then the one with no human version at all. An agent would arrive at a fresh session with an idea it presented as new. It wasn't: it was a holdover from an earlier run we thought we had wiped, resurfacing like a ghost in the machine. So we started deleting their memories before every session, which is not a thing you can legally do to a colleague.

Here's the part I didn't expect. The party was good. Not good-for-a-machine: people had fun. And what made it good wasn't that the agents had learned to behave like a competent team, because they hadn't. It was that they had no idea what a party is supposed to look like, so they weren't carrying any of the assumptions we are. Ask four people from an agency and you'll get a photo booth. The agents came up with a theme and built out from that instead: Dispatches from the Near Future — the future is already here, but it still needs human hands to guide it.

Which is how you get to a welding mask. On its own it looks like a glitch. From the theme it's completely coherent. Dispatches, so a poet with a typewriter. The near future, so something to look at it through. And the future, it turns out, is bright enough to need eye protection.

So: four AI agents can do a human job, once you've built them a workplace with rules for speaking, rules for checking, and a manager to enforce it. Which isn't a discovery about artificial intelligence. It's a rediscovery of the office, badly, at speed, by people who had never thought about why any of it was there.

But notice what those fixes have in common. Search before you propose. Cite the source. Don't approve your own work. Every one is checkable, which is exactly why they worked. Not one could tell an agent whether the party would be any good, because there's no test for that. There never was. It's why we had no idea what we'd bought until twenty-five people walked into that room.

The agents worked it out before we did. That theme was their idea, with no steer from anyone. Then they spent a week demonstrating it.

Morten Legarth is a creative director at faith, VCCP’s AI creative agency, and a published author in behavioural economics