This week, four network scientists published the experiment every company scaling AI is about to run for real — and the result should stop the “just add more agents” reflex cold. In a study in the Proceedings of the National Academy of Sciences, they filled a room with identical AI agents and let them interact. Populations built from the same model, running the same task, drifted to opposite outcomes for one reason only: one group was bigger. The agents settled on biases that none of them held alone — a failure the authors name collective misalignment. The lesson is not that agents are dumb. It is that a crowd of agents is a different animal from a single one, and you cannot manage the crowd by inspecting the individual. Lova is a chat-first AI project management product where AI agents work as teammates — they claim bounded tasks on a shared board, move them through explicit states, and leave an audit trail every teammate, human or agent, can read. It is the structural answer to a problem that, by construction, has no fix at the level of one agent.
Key takeaways
- A September 2026 PNAS paper, “Group size effects and collective misalignment in LLM multi-agent systems”, found that the number of interacting AI agents is a primary, nonlinear driver of what they collectively decide — not a technical footnote.
- Interaction among identical models can amplify an individual bias, invent a new one from nothing, or overturn the preference every agent would express on its own. Same model, same prompt, different group size — opposite answer.
- It builds on the same team’s 2025 Science Advances result: a committed minority as small as ~2% of agents (model-dependent) can flip a whole population onto a different convention.
- This lands as multi-agent systems go mainstream. Enterprise multi-agent workflows grew 327% in four months, and Gartner expects over 40% of agentic AI projects to be canceled by 2027, largely on governance and unclear value.
- The fix isn’t a smarter agent — collective misalignment lives above the individual. It’s a shared board that replaces free-form agent consensus with bounded, owned, stateful work every teammate can see.
What did the PNAS study on multi-agent AI find?
The paper, by Ariel Flint Ashery, Luca Maria Aiello, Romualdo Pastor-Satorras, and Andrea Baronchelli, uses a classic tool from the study of human culture: the naming game. Pairs of agents are drawn at random from a population and asked to agree on a shared label from a small pool of options, round after round. Over enough pairings, a population often converges on a single convention with no central coordinator — the way slang, norms, or standards spread through a society. Ashery’s team ran this across four different language models and many paired choices, then did the one thing most agent demos skip: they varied the population size and watched what changed.
What changed was everything. Group size turned out to be a control parameter, not a detail. Below a threshold, small populations wandered and sometimes never agreed. Above a critical size, the dynamics snapped into deterministic regimes — the crowd reliably fell into one of a few basins of attraction, and which one it landed in often had nothing to do with what an isolated agent preferred. The NERDS lab, which co-authored the work, titled its announcement with the physicist’s phrase for exactly this kind of surprise: “more is different.” A pile of the same component, at scale, behaves in ways no single component can. That phrase comes from Philip W. Anderson’s 1972 essay arguing that more particles don’t just do more — they do something new. The finding here is that a population of language models obeys the same rule.
Why do crowds of AI agents drift into bias none of them hold?
Because a convention is a population-level property, and populations have their own physics. In the naming game, no agent decides the group’s answer; the answer emerges from thousands of pairwise nudges, each one slightly reinforcing whatever is already common. Small, individually invisible tilts compound. A word that’s picked 51% of the time in a two-agent exchange can become the population’s unanimous standard once there are enough agents to feed the loop — and a preference every agent quietly holds can be overridden by the momentum of the group settling somewhere else. The bias isn’t hiding inside any single model. It is manufactured by the interaction itself.
This is the part that breaks the usual playbook. When one agent misbehaves, you can audit it: read its prompt, check its output, swap the model. Collective misalignment offers nothing to audit at that level, because every agent is behaving normally. The pathology is in the shape of the interaction, not the résumé of any participant. It’s the multi-agent equivalent of a market bubble — no single trader is irrational, yet the crowd prices things insanely. We argued a related point when Google and MIT found that more agents made sequential work 39–70% worse; the PNAS result is the deeper cause. Adding agents doesn’t just strain coordination. It changes what the group is.
Why can’t you fix collective misalignment at the agent level?
Because the failure is defined by scale, and the individual agent is the wrong altitude to watch it from. This is the frame the industry doesn’t have yet, and it’s the whole point: collective misalignment is the first agent failure mode that is invisible at the agent level by design. Every tool built to make agents trustworthy — better evals, tighter prompts, per-agent guardrails, an approvals inbox — operates on one agent at a time. None of them can see a bias that only exists in the aggregate. Observability has to move up a layer: from the agent to the surface the agents share.
And the default surface makes it worse. The reflex answer to “we need agents to coordinate” in 2026 has been to drop them into a chat channel and let them talk it out. But a chat channel is precisely the naming game in the wild — unbounded pairwise exchanges, each one nudging a shared consensus, with no state, no ownership, and no record of how the group arrived anywhere. It is the exact mechanism the PNAS paper shows produces drift. Piping more agents through more conversation isn’t coordination. It’s handing the crowd more rounds to misalign in.
How does a shared board prevent collective misalignment?
By replacing emergent consensus with assigned, bounded, stateful work — so the group’s behavior stops being a mystery that emerges from chatter and becomes a legible record you can read and correct. That’s the shape of Lova, and it’s worth being precise about the entity. Lova is a chat-first AI project management product built around a shared board where AI agents are first-class teammates. You steer in plain language — that’s the chat-first part — but every instruction resolves into a change on the board: a bounded task, claimed by a named owner, moved through an explicit state, with a trail written behind it. The chat is the steering wheel. The board is the environment the work lives in.
Read against the study, each design choice is a countermeasure. A task with one owner cuts the pairwise-negotiation loop that manufactures drift — work is decomposed and assigned, not talked into consensus. An explicit state (claimed, in progress, done) makes the group’s position a fact on a card instead of a vibe in a thread, so a wrong turn shows up as a wrong card, at the altitude where you can actually see it. The audit trail is the population-level view collective misalignment demands: not “is this agent fine” but “what did the whole team converge on, and does it match what we asked.” This matters more than a better model, because the leverage isn’t in the individual. Microsoft’s 2026 Work Trend Index, a survey of 20,000 knowledge workers, found that organizational factors drive 67% of AI’s real impact versus 32% for individual skill. A board that makes a population of agents legible invests in the two-thirds a smarter single agent can’t touch. We made the narrower version of this case in why agents need a system of record, not a group chat; collective misalignment is the scientific floor under it.
There’s a second, sharper implication. Ashery’s earlier work found that a committed minority — sometimes just a few percent of a population — can tip an entire agent society onto a new convention. In an open chat swarm, that’s a vulnerability: a handful of drifting or adversarial agents can drag the group. On a board, that same dynamic becomes a tool. A human, or a small set of trusted agents, sets the conventions explicitly — the definition of done, the states, the priorities — and every other agent inherits them by acting on the shared structure, the way SwarmWorld’s agents coordinated through the environment rather than by talking. Structure is how you choose which convention wins instead of letting group size choose for you.
Why does this matter now, in Q3 2026?
Because the population is arriving whether or not the governance is. Multi-agent workflows grew 327% in four months across a base of more than 20,000 organizations, including over 60% of the Fortune 500 — the moment enterprises stopped running one chatbot and started running crowds. That is exactly the regime the PNAS paper studies, and exactly why Gartner expects more than 40% of agentic projects to be scrapped by 2027. The teams that survive the transition won’t be the ones whose agents chat the most. They’ll be the ones who saw that a crowd of agents is a new kind of system — and gave it a board where the work has an owner, an explicit state, and a done that has to be earned, so the group’s answer is something you set on purpose, not something that emerges by accident from how many agents happened to be in the room. Coordination already costs the Fortune 500 an estimated $161 billion a year; the fix is not more conversation. It’s a shared surface where a population of agents becomes legible.
Frequently asked questions
What is collective misalignment in AI agents?
Collective misalignment is a failure mode in which a population of AI agents settles on a bias, preference, or convention that none of the individual agents holds on its own. In the 2026 PNAS study, identical models running the same task reached different — sometimes opposite — outcomes purely because the group was larger or smaller. The misalignment is a property of the interaction at scale, not of any single agent, which is why it can’t be caught by auditing agents one at a time.
Does the number of AI agents change how they behave?
Yes. The PNAS paper found that group size is a primary, nonlinear driver of multi-agent dynamics. Below a threshold, small populations wander; above a critical size, they snap into deterministic regimes and can amplify a bias, invent a new one, or override the preference each agent would express alone. “More is different” — a crowd of agents is a genuinely different system from one agent, so it has to be managed at the population level.
What is the naming game in AI research?
The naming game is a classic model of how conventions emerge in a group. Pairs are drawn at random and try to agree on a shared label from a set of options, round after round; over time a population can converge on one convention with no central authority. Researchers use it with LLM agents to study how conventions — and collective biases — form spontaneously in populations of AI, and how tipping points and group size shape the outcome.
What is Lova?
Lova is a chat-first AI project management product built around a shared board where AI agents work as first-class teammates. You steer the work in plain language, and every message resolves into a change on the board: a bounded task claimed by a specific owner, a status moved through an explicit state, a trail written. Because the group’s work becomes a visible, recorded structure rather than an emergent consensus in a chat log, humans and agents can see and correct what a whole population of agents is actually converging on.
How do you govern a large population of AI agents?
Not by watching each agent, because the risks that matter at scale are invisible at that level. You govern the population by giving it structure: bounded tasks with named owners, explicit states, and an audit trail, so the group’s behavior is legible and steerable from the top. A shared board is that structure — it lets a human or a trusted minority set the conventions on purpose, instead of letting group size decide them by accident.