What AI Teams Teach Us About Human Ones
Interested in trying ChatCEO, our new AI platform for CEOs? If you lead a company of more than 50 employees, you are eligible to be a free early user. Message me to get started.
A couple of weeks ago, I wrote this piece about how AI agents need good management, just like people:
As we build our AI system for CEOs, I was struck by how various agents needed the same things human workers do:
Clear instructions
Feedback
Ability to coordinate
etc., etc.
This week, I read a study that complicated that comparison significantly. The similarities are still there, and AI agents still need managers. But in some ways, collaboration between AI agents is pretty different from how humans work.
I think it’s critical to know the difference.
The Study: How AI Agents Work Together
A researcher at the Moscow Institute of Physics and Technology ran more than 25,000 task simulations across eight different AI models. Agents were put in groups ranging from 4 to 256, organized in four different coordination structures.
The central question was: Does the coordination structure between agents affect the quality of the output?
The answer was an unambiguous YES. The coordination mechanism mattered just as much, if not more, than which model you chose.
(This is itself a similarity to the human world. A team of B-players working at a business with a strong management system will usually dominate a team of A-players within a broken organization.)
Testing 4 Coordination Structures
The study tested four main coordination approaches. As you read, guess which one you think won:
Coordinator - one agent analyzes the task, assigns roles to other agents, and they execute in parallel
Shared - every agent has access to a shared memory of past work and makes all decisions simultaneously and independently.
Sequential - agents work in a fixed order, with each seeing the completed output that came before and deciding what to do next.
Broadcast - agents first signal their intentions to each other simultaneously, then make final decisions informed by what everyone else said they were going to do.
Which do you think won?
It was Sequential, by a landslide.
Sequential outperformed the centralized Coordinator structure by 14%, and outperformed the fully autonomous Shared structure by 44%, an effect size the researchers describe as enormous.
The constraint of a fixed sequence unlocked a degree of spontaneous alignment that would have been hard to achieve by design. The reason comes down to what each agent actually sees. In the Sequential structure, every agent observed what its predecessors actually produced on the current task and used that to conclude what the next best action was (and sometimes that meant they would not contribute at all to the next step).
The researchers describe it as a sports draft. Each team makes its pick knowing exactly who has already been selected, and naturally fills the gaps without anyone running the process from the top.
Does Any of This Apply to Managing People?
That’s really important data for us to have as we assemble teams of agents going forward.
But it’s also interesting, again, to think about the similarities to the management of humans. Below are some of my core beliefs about management within organizations, with thoughts on whether - in light of this paper - they hold true for AI agents too.
✅ Shared mission is the foundation.
When agents were given a clear mission and the freedom to determine how to contribute to it, they achieved perfect mission relevance scores. The researchers’ practical conclusion was that agents need a clear mission and the right structure. That maps onto what I’ve always argued about human teams. Clarity of purpose is where it all begins. If you haven’t made your 1-Page Strategic Plan and shared with your team, do it today.
✅ “Autonomy within structure” beats micromanagement and anarchy.
Neither the tightly controlled Coordinator nor the unconstrained Shared protocol performed as well as Sequential. The sweet spot (for people and apparently for AI) is freedom to contribute within a structure light enough not to get in the way. It’s the lessons of the Prussian Army and mission command still holding relevance today. For a primer on this concept, Bungay’s Art of Action is the gold standard.
✅ Transparency is powerful.
I have long argued that all employee goals in an organization should be visible to all other employees. The Sequential protocol’s advantage is built almost entirely on this principle. Each agent sees what every predecessor actually produced before it begins its own work. The researchers tested every other approach (shared intentions, shared history, a coordinator’s plan) and none performed as well as simply showing each participant the completed work of those who came before them.
✅ Throwing more agents at a problem doesn't help.
Growing a business does require growing a team. But that's different from the impulse to accelerate a lagging project by flooding it with new resources, a trap I've written about before. The study bears this out at the agent level with striking clarity: scaling from 64 to 256 agents produced no measurable improvement in output quality, at 4.6 times the cost. Meanwhile, the quality gap between the best and worst AI models tested reached 174%. The right model matters far more than the number of agents running it. For AI teams as for human ones, adding bodies is rarely the answer to a performance problem.
❌ AI agents don’t need the kind of hierarchy that human organizations do.
I believe in traditional management structure for a reason. A CEO needs an executive team covering the six core functions of the business. Those executives need managers beneath them, each with a reasonable span of control, somewhere between four and ten people. That structure exists because humans need context, coaching, and someone accountable for their growth.
AI agents need almost none of that. Left to their own devices in this study, the agents organized into at most two layers of structure regardless of how many agents were involved. They created just enough coordination to complete the task and no more. Holacracy and “flat” orgs have usually failed in the human realm, but it might be what agents need.
❌ Pre-assigned roles may actually limit AI performance.
The researchers argue that assigning fixed roles to AI agents imposes human limitations onto systems that don’t share those limitations. Humans usually do better with role clarity: what skills are needed, what KPIs determine success in the role, what will I be held accountable for, etc. But agents in the study performed better when they could self-select their function based on what the task needed, rather than executing a role someone had defined for them in advance. There’s gray area here: no job description should become a set of shackles for a human, but it appears that a foundational role description is more of a human need than an AI one.
Bottom line: AI coordinates like humans do in some ways, but not in others. And structure matters, regardless of what’s inside the system.