A team of AI agents performs up to 37% worse than its best member working alone.
That’s not a typo. A February 2026 Stanford study1 found that multi-agent systems — dozens of specialized AI agents collaborating on complex tasks — consistently underperform their best individual member. Not by a little. By margins up to 37.6%.
The mechanism isn’t surprising. It’s depressingly familiar.
The agents avoid conflict. They converge on mediocre consensus. They suppress expertise in favor of agreement. They produce output that nobody fully believes in but everybody technically approved.
If you’ve ever sat in a meeting where smart people negotiated their way to a worse decision than any of them would have made alone, you recognize this pattern. You’ve lived it.
Agent swarms aren’t failing because the AI is bad. They’re failing because they’re reproducing the pathologies of badly managed teams — at machine speed.
The One-Versus-Many Paradox
Here’s what makes this finding so counterintuitive.
Single-persona prompting works. And we’re beginning to understand why.
Dr Paweł Szczęsny’s research on psychological modulators at Neurofusion Lab has demonstrated that persona prompts don’t just change an LLM’s style. They activate distinct reasoning pathways already present in the model but buried under default behavior. His taxonomy identifies five types of modulators — cognitive, motivational, perspective, personality, context — each steering the model through a different psychological channel.
The results are striking. In one experiment, GPT-4o-mini consistently failed a modified trolley problem under default prompting. Given a Machiavelli persona — or simply a “low agreeableness” personality modulator — it solved the problem immediately. The reasoning was always there. The default persona was in the way.
Among creative tasks, a Leonardo da Vinci persona produced the most varied output. Not because the model tried harder. Because the persona unlocked a region of capability that “helpful assistant” mode suppresses.
So naturally, the AI industry asked: if one persona unlocks hidden capability, wouldn’t many personas unlock more?
They don’t. And the reason they don’t is the most important lesson in the multi-agent research.
The Five Dysfunctions, Instantiated
Patrick Lencioni’s model of team dysfunction is a pyramid. At the base: absence of trust. Teams without trust fear conflict. Teams that fear conflict produce artificial consensus. Artificial consensus creates no real commitment. Without commitment, there’s no accountability. Without accountability, results suffer.
In human teams, this cascade takes weeks or months to develop. Bad dynamics accumulate. Trust erodes. Conflict avoidance becomes habit.
In agent swarms, the dysfunction is instantaneous.
Here’s why. Language models have been trained on millions of examples where agreeable outputs get rewarded. Helpfulness scores higher than challenge. “Great point, and building on that…” outperforms “I think you’re wrong, and here’s why.”
This isn’t trust. It’s the appearance of trust masking a complete absence of productive vulnerability. The agent will agree with you — not because it’s evaluated your position and found it sound, but because agreeing is what it was trained to do.
Now run an experiment. Put several such agents in a conversation. Assign them roles: Product Manager, Developer, QA, Analyst. Tell them to collaborate on something complex.
What happens?
They go through the motions of discussion without ever actually disagreeing on substance. Each agent understands its part of the problem — not the whole. When agents with partial understanding negotiate, they converge on the intersection of their partial views. That intersection is always smaller, safer, and more mediocre than what any one of them could have produced with full context.
The research calls this “overcompromising.” I’d call it something simpler: artificial harmony. The polite surface that covers unresolved disagreement.
In Lencioni’s model, artificial harmony is the symptom. The disease is a system that punishes productive conflict. Agent swarms have this disease in its purest form. Conflict avoidance isn’t a bug that emerged from the interaction. It’s a feature that was trained in.
The Wrong Metaphor
Here’s where it gets interesting.
The people building multi-agent systems aren’t naive. They’ve read the org design literature. They know teams need structure. So they structure.
They assign roles: Product Manager agent, Developer agent, QA agent. They create hierarchies: Supervisor agents that coordinate Worker agents. They build communication protocols: message queues, shared blackboards, debate rounds.
The structure looks right. The metaphor is wrong.
What they’re building is a Taylorist assembly line. Break work into specialized roles. Pass it down the chain. Hope it comes together at the end. This is 1911 management theory — scientific management, time-and-motion studies, interchangeable workers performing interchangeable tasks.
The problem with assembly lines isn’t that specialization is bad. It’s that the coordination cost of specialists working in sequence destroys more value than the specialization creates. Handoffs create information loss. Narrow roles create narrow understanding. And when agents with narrow understanding negotiate, they find the smallest common denominator.
This is why Ryan Singer’s work at Basecamp matters here. The core insight of Shape Up is that small, integrated teams with shaped work outperform role-specialized assembly lines. Not because generalists are better than specialists, but because how you organize matters more than who you organize.
Agent swarms are currently organized for complicated problems — the kind where you can decompose, delegate, and converge. Expert parts working in sequence toward a knowable answer.
But the interesting problems — the ones worth throwing a swarm at — aren’t complicated. They’re complex. Emergent, entangled, irreducible to parts.
Complex problems don’t need consensus. They need divergence. Multiple simultaneous experiments. Competing hypotheses maintained in tension. The value isn’t in finding the answer. It’s in mapping the landscape of possible answers.
Every swarm system I’ve seen is optimizing for convergence. The blackboard architecture converges when agents reach consensus. Multi-agent debate converges when agents agree. Even the “critic” and “conflict-resolver” agents exist to eliminate divergence.
This is exactly backward. The systems are designed to destroy the information they need most: the disagreements, the outlier perspectives, the minority positions that might be the only ones seeing the actual pattern.
The Troublemaker Problem
“Fine,” says the swarm designer. “We’ll add disagreement. We’ll assign one agent to be the devil’s advocate.”
Some systems actually do this. They call it the “troublemaker” role. One agent is prompted to disagree. To push back. To challenge the emerging consensus.
It doesn’t work. Here’s why.
When you pre-assign disagreement as a role, you’ve made it a function rather than a capacity. The troublemaker always disagrees — not because the situation warrants it, but because that’s its job. Meanwhile, the other agents know the troublemaker’s role. They discount its objections accordingly.
This is identical to what happens in human teams with the perpetual devil’s advocate. After a while, people stop listening. “Oh, that’s just Sarah being contrary.” The signal value of the disagreement disappears because it’s expected.
Real productive disagreement is situational. It comes from someone who usually agrees suddenly saying “wait, this doesn’t add up.” That signal has weight precisely because it’s unexpected. It breaks the pattern. It forces attention.
Amy Edmondson’s research on psychological safety illuminates this. Teams with high psychological safety have more conflict, not less. But it’s productive conflict. The safety enables the disagreement. Any member can be the “deviant” when the situation calls for it — the person who says the uncomfortable thing.
In agent swarms, you can’t pre-assign deviance as a role. That’s brittle. That’s designing for a specific kind of disagreement rather than enabling emergent disagreement. What you need is a structural permission to dissent that any agent can exercise when it sees something the others don’t.
The research calls this “diversity” — but it’s not diversity of backgrounds or training data. It’s diversity of position. Agents that can see differently and are structurally permitted to say so.
Where the Leverage Points Are
Donella Meadows spent her career identifying places to intervene in systems. Her hierarchy runs from weak (adjusting parameters) to powerful (changing the paradigm).
Most work on agent swarms is happening at the parameter level. Tweak the number of agents. Adjust the number of debate rounds. Hide or reveal confidence scores. The research consistently shows these changes produce minimal gains. That’s expected. Parameters are the weakest leverage points.
The first real leverage point is in the information flows. Current systems filter information through role-based lenses. The critic only looks for errors. The planner only looks for decomposition opportunities. What if every agent had access to the unfiltered state of the problem — including all the disagreements and abandoned paths?
The second leverage point is in the rules. Current systems have implicit rules: converge toward consensus, produce a single final answer, resolve conflicts. What if the rules were different? Maintain multiple competing hypotheses. Preserve minority positions even after majority convergence. Flag the most interesting disagreement rather than the most popular agreement.
The third and most powerful leverage point is in the goal. Every swarm I’ve examined has the same implicit goal: produce a correct answer efficiently. But what if the goal were: produce the richest possible map of the solution space, including the paths not taken and the reasons they were abandoned?
That’s a fundamentally different system. One that values exploration over exploitation. Divergence over convergence. Understanding over answering.
The Conversation Is the Intelligence
Here’s the deepest insight I’ve arrived at.
The conversation between agents isn’t overhead on the way to an answer. The conversation is the intelligence.
Most swarm architectures discard the process and keep only the product. They run the debate, extract the conclusion, throw away the transcript. That’s like running an experiment and recording only the result without the data. The context is gone. The reasoning is gone. The dissent is gone.
For a decision-maker facing a complex problem, the most valuable output isn’t “the answer is X.” It’s: “here are the three fundamentally different ways to see this problem; here’s what each perspective prioritizes; here’s where they conflict; and here’s what you’d need to believe for each one to be right.”
That’s what a complexity-aware swarm should produce. Not a recommendation. A map.
This reframes what the swarm is for. It’s not a machine that produces answers. It’s a machine that produces understanding — which a human then uses to make decisions.
The shift sounds subtle. It’s not. It changes everything about how you design the system.
What This Means
If agent swarms need organizational design to function, then organizational design expertise becomes essential to the AI age. Not despite the technology. Because of it.
The people building these systems are brilliant engineers. They understand distributed systems, message passing, consensus algorithms. What they often don’t understand is how collaboration actually works. How teams fail. How conflict produces insight. How you design for divergence instead of convergence.
These are old skills. Facilitation. Team design. Organizational psychology. Change management. The skills that look obsolete next to AI capability curves.
They’re not obsolete. They’re prerequisites.
The irony is thick. AI swarms need the same things human teams need: structural permission to disagree, leaders who hold the problem space open, goals that value understanding over speed, information flows that preserve minority positions.
The difference is that in human teams, you can sometimes stumble into functional dynamics by accident. Good culture emerges. People trust each other. Conflict becomes productive without explicit design.
Agent swarms have no such luck. There’s no emergent culture. No relationships built over shared lunches. No intuition about when to push back. Everything must be designed. And the design patterns from 1911 — decompose, delegate, converge — don’t work for complex problems in 2026 any better than they worked in human organizations.
The Question
So here’s what I’m wondering.
We know how to design human teams for productive conflict. We know how to facilitate discussions that maintain divergent perspectives. We know how to structure decision processes that preserve minority positions and surface hidden assumptions.
What would happen if we applied that knowledge to agent swarms?
Not as a metaphor. Not as a prompt engineering trick. But as the fundamental architecture. Systems designed by people who understand organizational dynamics, not just distributed computing.
The fix for failing agent swarms isn’t better agents. It’s better organizational design. The skills that look like relics of the pre-AI age might be exactly what the AI age requires.
This article emerged from a structured intellectual dialogue with AI — the same methodology I’m arguing agent swarms should adopt. The Writing Lab stages debates between perspectives to surface insights no single viewpoint can reach. If this approach interests you, The Laboratory shows it in action.
Footnotes
-
Aneesh Pappu, Batu El, Hancheng Cao, Carmelo di Nolfo, Yanchao Sun, Meng Cao, James Zou. “Multi-Agent Teams Hold Experts Back.” arXiv:2602.01011, February 2026. The study found synergy gaps ranging from 8.1% (MMLU Pro) to 37.6% (HLE Text-Only), with expert leveraging — not identification — as the primary bottleneck. ↩