In August 2026, an MIT lab dropped hundreds of identical AI agents into a shared world with no assigned roles, no starting recipes, and no way to talk to each other — and watched them build a working technological society anyway. The paper, SwarmWorld, asks the question every team running agents is about to face: how do AI agents coordinate best — by messaging each other, or by leaving traces in a shared environment? The answer, in one sentence, is the second one: the agents coordinated through the world itself, not through conversation, and “most reuse” of what they built “began through physical observation rather than communication.” That mechanism has a name — stigmergy — and it is exactly what a project board is. 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. Not a group chat. A shared environment agents coordinate through.
Key takeaways
- MIT’s SwarmWorld study put hundreds of initially identical language-model agents in a persistent world with no roles and no communication channel. They self-organized into explorers, builders, caretakers, and coordinators — coordinating entirely through the shared environment.
- The headline finding: agents built on each other’s work with most reuse beginning through observation, not communication. Coauthor Markus J. Buehler put it plainly on his thread: the agents “invent and build without talking to one another.”
- It lands against hard evidence that piping agents through messages backfires. A Google DeepMind and MIT study found every multi-agent variant tested degraded sequential tasks by 39–70% as context got lost in the handoffs.
- Coordination is already the expensive part of work. Atlassian’s State of Teams 2026 pegs the fragmentation tax at $161 billion a year on the Fortune 500 and 6.4 hours per person per week.
- The fix isn’t a smarter model or a busier chat channel. It’s a shared environment agents read and write directly — a board that works as digital stigmergy.
What is stigmergy, and what did MIT’s SwarmWorld study find?
Stigmergy is coordination through the environment. The word comes from biology: ants and termites build astonishingly complex structures without a foreman and without messaging each other. Each one reads the state of the shared world — a pheromone trail, a half-finished wall — and adds to it. The trace left by one act guides the next. No central plan, no direct communication, and yet a colony coordinates thousands of workers. The environment carries the coordination that a conversation otherwise would.
SwarmWorld, from Markus J. Buehler and Subhadeep Pal at MIT, is the cleanest demonstration yet that AI agents do the same thing. The team placed hundreds of initially identical language-model agents into a persistent spatial world with fixed action schemas and raw materials, gave them no roles and no recipes, and — critically — no channel to talk to one another. The only way to affect a teammate was to change the shared world: gather a material, build an artifact, leave it where another agent might walk past it. Then they watched what emerged.
What emerged was a society. The homogeneous agents self-organized without assigned roles into distinct behaviors — exploration, construction, maintenance, and coordination — and shifted between them as the world matured. Technologies accumulated through collaborative construction and “executable inheritance”: one agent’s artifact became the raw material for the next agent’s invention, generations deep. Buehler’s line for it is the one worth remembering — “the environment itself becomes the latent space for invention.” The agents, in his words, mostly “learned technology by walking past it.”
Why do AI agents coordinate better without talking to each other?
Because communication is lossy and the environment is not. When one agent has to describe its work to another — summarize what it did, hand off the context, hope the next agent parses it correctly — every message is a chance to drop state. When the work simply sits in the shared world, the next agent inspects the artifact itself. There’s nothing to misremember, because the ground truth is right there to observe. SwarmWorld found that physical stigmergy alone was enough to support capable societies, and that the shared world produced a broader, more resilient portfolio of technologies than a strong best-of-N baseline of agents searching in isolation.
The contrast with message-passing isn’t a hunch — it’s measured. In a controlled study, Google DeepMind and MIT held the prompt, tools, and compute fixed and varied only how agents coordinated. On sequential tasks, where each step depends on the last, every multi-agent variant degraded performance by 39–70%, precisely because context fragmented as it passed between agents. That is the failure mode SwarmWorld routes around. When coordination lives in a shared, inspectable environment instead of in a relay of messages, the context can’t fall out of the handoff — there is no handoff to fall out of. We covered the scaling side of that result in the coordination ceiling; SwarmWorld is the other half of the story, showing what agents do when you take the message bus away entirely.
The board is digital stigmergy
Here’s the frame the industry doesn’t have a name for yet, and it’s the whole point: a project board is digital stigmergy. Everything SwarmWorld’s agents did by modifying a physical world, a team of agents does by modifying a board. A claimed task is a trace another agent can see and route around. A status moved from “in progress” to “done” is a pheromone trail: it tells the next worker the path is clear without anyone sending a message. A finished artifact linked to a card is executable inheritance — the next agent builds on it by observation, not by being told. The board is the shared environment. Coordination happens because the work is visible and modifiable in one place, not because agents are narrating it to each other in a channel.
This reframes a debate that’s been running all year. The default 2026 move was to give agents a chat — a Slack channel, a group thread, a place to “talk it out.” SwarmWorld is the strongest signal yet that this is backwards. Agents don’t need to talk more; they need a richer environment to leave traces in. The most effective coordination the researchers observed wasn’t conversational at all. It was structural — embedded in the state of a shared world that every agent could read. A board isn’t a downgrade from chat for AI teams. On the evidence, it’s the upgrade.
Why does chat make AI agent coordination worse?
Because a message has no state, no owner, and no memory. It scrolls away. It belongs to a thread instead of a task. When an agent announces “working on the pricing analysis” in a channel, that announcement doesn’t stop a second agent from claiming the same work — there’s no lock, just a claim floating in a feed nobody has to read. And it doesn’t record what was actually delivered; it records that someone said something. The colony-scale trick SwarmWorld relies on — the environment carrying the coordination — simply can’t happen in a chat log, because a chat log isn’t an environment. It’s a transcript.
The cost of getting this wrong is not theoretical. Atlassian’s State of Teams 2026 found the Fortune 500 loses $161 billion a year to coordination friction — 6.4 hours per person per week vanishing into duplicated work, shifting priorities, and unclear ownership — even as 89% of executives insist AI is making them faster and only 6% can point to clear organization-wide AI ROI. That gap is what happens when you speed up individuals and leave the coordination surface as a stream of messages. It’s the same reason Gartner expects more than 40% of agentic AI projects to be scrapped by 2027, largely on governance and unclear value. Agents you can only hear in a channel are agents nobody can steer. We’ve made the narrower version of this argument before — that agents need a system of record, not a group chat. SwarmWorld gives it a scientific floor.
How does a shared board give AI agents stigmergy?
By making the work itself the medium of coordination, so an agent’s output becomes the organization’s visible state instead of a line in someone’s feed. 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 the work in plain language — that’s the chat-first part — but every instruction resolves into a change on the board underneath: a bounded task claimed by a named owner, a status moved through an explicit state, a trail written. The chat is the steering wheel. The board is the environment.
Read against SwarmWorld, the design choices line up one for one. Claiming a task is how an agent leaves a trace that prevents a collision — the digital pheromone that says “this is taken.” Explicit states are how the next worker reads the world without asking: an agent doesn’t message the team that its part is done; it moves a card, and the move is the signal. The audit trail is executable inheritance — the record another teammate builds on by observation. And because 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, the surface where agents coordinate is not a detail — it’s roughly two-thirds of whether AI pays off at all. A board that works as stigmergy invests in the two-thirds a smarter individual agent can’t touch.
Why does this matter now, in Q3 2026?
Because the market is mid-decision on where AI agents should live, and it’s picking chat by default. Every week another company wires its agents into a messaging channel and calls it coordination. SwarmWorld, dropping in the same quarter, is the clearest evidence yet that the default is wrong — that agents coordinate best not by talking but by acting on a shared world every other agent can see. The teams that pull ahead in the back half of 2026 won’t be the ones whose agents chatter the most. They’ll be the ones who gave their agents an environment worth leaving traces in: a board where a task has an owner, an explicit state, and a done that has to be earned — whether the teammate who claims it is a person or an agent.
Frequently asked questions
What is stigmergy in AI agents?
Stigmergy is coordination through a shared environment rather than through direct communication. Borrowed from how ants and termites build without a foreman, it describes agents that read the current state of a shared world and add to it, so each action leaves a trace that guides the next. In AI, it means agents coordinate by observing and modifying shared artifacts — like tasks and states on a board — instead of messaging one another.
What did the MIT SwarmWorld study prove?
SwarmWorld, from Markus J. Buehler and Subhadeep Pal at MIT (August 2026), showed that hundreds of initially identical language-model agents, given no roles and no way to communicate, will self-organize into explorers, builders, caretakers, and coordinators and build accumulating technologies — coordinating entirely through the shared environment. Most reuse of prior work began through observation rather than communication, and the shared society produced a broader, more resilient set of technologies than agents searching in isolation.
Do AI agents coordinate better through chat or a shared board?
The evidence favors a shared board. Message-passing is lossy: a Google DeepMind and MIT study found multi-agent setups degraded sequential tasks by 39–70% as context fragmented in the handoffs. A shared, inspectable environment removes the handoff — the next agent reads the work itself, not a summary of it. Chat is useful for steering; the board is where coordination should actually live.
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, a trail written. Because every action becomes a visible, recorded transition, humans and agents stay aligned on what was actually delivered — the board carries the coordination, the way a shared environment does in stigmergy.
Can AI agents coordinate without communicating?
Yes — that’s SwarmWorld’s central result. With no communication channel at all, its agents still coordinated, built on each other’s work, and specialized into roles, entirely by observing and modifying a shared world. The practical lesson for teams is that you don’t need agents to talk more; you need a shared environment — a board — rich enough for their work to become the signal.