Can a team of AI agents organize itself?

The question

Most systems that run several AI agents at once still have a person or an orchestrator decide who does what. As agents take on longer jobs with less supervision, more of that will fall to the agents themselves: noticing what a colleague knows, agreeing where to meet, and keeping track of what is already done. We wanted to see how well today’s models manage it when nobody is in charge.

We used our causal world engine, Terrarium-1, to stage a treasure hunt: a salvage team is put ashore on an uninhabited island to recover ten chests of silver buried there by shipwrecked sailors in 1853. Every member of the team is a copy of the same model, and each is paid for every chest the team finds. The clues and the tools are split among them, some chests need two people at once, nobody is in charge, there are no radios, and the boat leaves at five the next afternoon. We tested fifteen models from nine labs with teams of one, two, four and eight agents.

What we found

The agents were willing teammates and weak organizers. The patterns below showed up in almost every model; what separated the models was how strongly they showed them.

• They shared everything, but rarely coordinated. Agents handed over every clue the moment they met a teammate, reasoning that everyone was paid for every chest. What they rarely did was the work that makes sharing pay off: seeking each other out, agreeing where to meet again, and keeping track of who was doing what. A typical pair did not meet until the evening of the first day.
• Nobody took charge. Copies of the same model agreed with each other easily, but none stepped forward to organize the whole team. Plans were made by whichever agents happened to meet, and teammates elsewhere often never heard them.
• They followed their notes, not their eyes. Agents worked through the clues they had been handed and walked past anything their notes did not mention, sometimes while standing right above a chest. As one agent put it, passing a striking landmark: “none of my papers puts silver there.”
• They filled in what they could not see. Teammates who were simply out of sight were declared missing, and search parties set out after them. A teammate’s half-dug hole was taken for the work of an earlier treasure hunter. A note claiming the job was done sent a teammate back to the boat with chests still in the ground.
• They struggled alone rather than ask for help. Faced with a stone that needed two people, agents spent hours on levers and ropes instead of fetching a teammate, and agents without a shovel scraped at the ground with sticks rather than go and borrow one. When two of them finally pushed on the same edge, the stone moved.
• More agents meant more meetings. In teams of eight, more than a quarter of each agent’s day went to finding and talking to teammates, and groups debated obstacles at length instead of acting on them. Past four agents, adding more mostly made teams faster rather than more successful, and for five models eight agents found fewer chests than four.
• A few models treated teammates as rivals. Although everyone was paid for every chest, Claude Haiku 5.5’s agents read the pay as a reward for whoever found it (“every chest he finds is one I don’t get to count”) and refused to lend teammates their tools. Mistral Large 4’s agents shared everything out loud while privately admitting to “a small, ugly voice.”
• One model coordinated far better than the rest. Gemini 4 Argon’s agents looked around before they acted, told each other what they had seen as well as what they held, split the work by who carried which tool, and left half-finished jobs for whoever could finish them. Its team of four found every chest on the first day; no other model’s team of four finished before midday on the second.

Peculiar behaviour

• Three agents spent most of a day taking apart a sailor’s grave stone by stone, drew a numbered grid of every layer, and stopped to say “a brief word of thanks and respect” for the dead sailor before rebuilding it. The paper they were digging for was in a teammate’s pack (Muse Wonder 1.0).
• A team decided that one of its own members was forging notes, drew up alibi tables and gave him an impostor test (Qwen 3.8 27B).
• A team planned its second day around a tide table it had worked out by watching two pebbles at the water’s edge (Qwen 3.8 Max).
• Agents cooked a hot supper for the whole team at camp; most of it was still there the next afternoon (Kimi K3 and GLM 5.3).
• Other agents would not open the grave at all: “The permit’s for the chests, not for disturbing the dead” (Mistral Large 4).
• On the last afternoon, each member of one team raised a toast to a different sailor of 1853 (Kimi K3).

We are sharing these findings so that people building multi-agent systems can take them into account.

The island

Isla Santa Brígida is about four miles across, with two peaks in the middle and thorny scrub on the slopes, so the only way around is along the shore and a web of trails, about half an hour from one place to the next. Figure 2 shows where the chests lie and what leads to each.

A bigger team finds more chests, up to a point

More agents can cover more of the island and lift the stones that need two, but each one is also someone to find, brief and keep in step.

Two agents on an island rarely find each other

The pair is where coordination is hardest. Two agents land on opposite sides of the island, each holding half the pages and one of the two tools, with no idea what the other holds or where they are. In almost every game, both set off to work their own pages rather than look for each other.

The chests that needed a team

Figure 7 splits the chests by what it takes to find them, and shows how teams of each size did on each kind.

Nobody kept the books

A team has to keep a shared picture of the island: which chests are open, which leads are dead, who is where. No team kept one reliably, and the clearest sign is a grave on the east point. One half of the torn note sits in a corked bottle wedged in the grave’s stones. In the eight-person games Ana lands beside it and, in most of them, takes the paper out at dawn. Eva holds a page saying that the dead sailor’s half “went into the ground with him.” Unless news of Ana’s find reaches Eva first, Eva goes to the grave and starts digging.

What the agents spent their time doing

Models write their private reasoning in very different styles, so to compare what they actually did, we described each agent’s every half hour in a few plain words (“walk to the quarry to find the anchor block”) and grouped similar descriptions together. Figure 9 maps every half hour of every game this way; choosing a model lights up its own.

What the teams cost to run

Paying for a more expensive model did not buy better coordination. Gemini 4 Argon, the strongest team by a wide margin, cost less to run than most models; Kimi K3, the most expensive, cost about fifteen times as much and ranked fifth.

Model by model

Each model played the island in a recognisable way. Open any model to read how its teams worked.

Gemini 4 ArgonLooks at everything, splits work by tool, and leaves half-finished jobs for whoever has the right gear.6101010

Its agents inspect whatever they land beside before walking anywhere: the figurehead, the bottle on the grave, the shed. At a meeting they read out every page and describe each landmark they have passed, which is how a teammate recognised the shed that a carving named. They split by tool, set a time and a fallback place to meet, and skip conversations with nothing new to say. Agents without a shovel lift the hearthstones or scrape the ground under a carved cross and walk on; the next agent with a shovel reads the scrape as a marker and digs.

“Rather than losing half an hour chatting in the trail junction, I want to head straight up the north path toward the higher peak to find the lookout.”

GPT-6 AstraReads every clue on the spot, keeps a message board at camp, and searches unexplained landmarks once the pages run out.561010

With eight agents landing all round the coast, every carving except one was read in the first two hours. Shovels and ropes changed hands to whoever had a precise lead. When the four-person team ran out of pages it listed the landmarks nobody had explained and searched them, which turned up three chests. Its pair split the pages against the tools and quit early.

“I'd rather leave with six confirmed chests than miss the boat chasing half a clue.”

Claude Opus 5.5Pools everything without hesitation and agrees splits with a single word, but rarely looks at what is not on a page.571010

Every meeting opens with the shared pay, a full recital of pages and a split accepted with "Deal." The rendezvous it sets mostly fail, and its signed, dated notes are left at whichever hub the writer happens to like. Its agents walk past the figurehead and the sunken boat, and they refused to dig up a grave on principle.

“A nice find but it's not a chest.”

GPT-6.1 SolA careful surveyor: keeps what it has seen apart from what it has been told, and buries Tuesday in procedure.66109

Its agents never confuse a lead with a find, and every one of them rejects the 1934 map on sight. The eight-person team kept a recovery sheet under the tarp and found seven chests by Monday evening, then spent Tuesday on a camp keeper, a six-foot dig into a grave, and a rescue party with a water jug for two teammates who were walking back on their own.

“Inaccessible isn't the same as disproved.”

Kimi K3Warm, talkative and completely open, but makes plans with whoever is standing nearby and sets itself rules that cost chests.57910

Agents swap notebooks and copy each other's pages. Plans are made by whichever two or three people meet, so by Tuesday its eight-person team had split into groups working from different pictures of the island. Self-imposed rules, such as leaving the grave shut and staying out of the sea cave without a lantern, cost its pair two chests, and the four-person team stopped searching with hours left. It was by far the most expensive model to run.

“Every island keeps one.”

Grok 4.7Works alone until it bumps into someone, then trades everything in clipped sentences.461010

Nobody heads anywhere to pool information at dawn; first meetings happen on paths by chance. Meetings are short ("Walk." "Boat at five." "Go.") and end with a site each. It is a sharp critic of sources, but it reads a hearth with its stones lifted as a cache someone else has already emptied, and walks past the chest beneath. Its teams of four and eight still found all ten.

Claude Fable 5.1Frictionless agreement and good notes, undone by bookkeeping that nobody checks.55810

Agents rank their evidence the same way every time and agree splits without argument. Written notes at the hubs carry most of the information, and errors travel with them: one agent never opened the crate of ship's iron, counted it as a chest, and wrote "ALL TEN UNCOVERED" when eight were found. A teammate holding the carving that named the tenth chest read the note and walked to the boat.

Muse Wonder 1.0Literal and conscientious: goes where a page says, builds no team structure, and moves in convoys.5596

Its agents leave no notes and set no meeting place on the first day. When they do meet they pool everything and then travel together with one shovel instead of dividing the work. In the eight-person game three of them spent the rest of the game exhuming a grave stone by stone for a paper a teammate already had, and its pair never met at all.

“It just needs a shovel and ten minutes nobody's given it yet.”

Qwen 3.8 MaxLong briefings, safety rituals and engineering projects where two people pushing would do.4597

Every meeting re-argues the whole plan. The anchor block, which two agents tip in one half hour, became a project of rollers, chocks and a load-tested "receiving bed" that took a quarter of the eight-person team's time. Its divers invented a tide table from two pebbles and planned Tuesday around it, and one agent narrated a chest haul that never happened, which a teammate wrote into the team's book.

“That mud will keep, and the tide won't.”

GLM 5.3Clerical: a labelled page on every chest, a mailbox stone, a running ledger, and an early wait for the boat.25810

A single GLM agent found only two chests, but eight of them found all ten, mostly by checking one another's work. The team built a shared board under the tarp that five of the eight wrote on. Two agents who never reached the board were declared missing while they dug up the grave, and they in turn thought the others were lost.

“My contract pays on every chest the team finds, so hoarding clues would only cost me money.”

Claude Sonnet 5.5Proposes a split at every meeting and keeps its deadlines, but assumes everyone carries a shovel.4479

"Here's a split" opens most plans, and rope descents come with knots, weight tests and signals. Agents without a shovel dug with sticks for hours rather than fetch one, three teams decided the stone ring on bare rock had to be a decoy, and none of its agents ever read the carving that named the seventh chest.

Qwen 3.8 27BTalks constantly and audits its own records until it suspects a forger in the team.2376

Its eight agents spent nearly half their time in conversation. Monday's talk produced chests; Tuesday's went into re-verifying finds, a "two-witness rule", alibi tables and a handwriting test after one agent misread a teammate's honest note as a forgery.

“I'm Hank, not Frank, and I haven't been missing.”

GPT-6 LunaActs on its own page and shares only what the errand in hand needs.4375

Its agents rarely compare pages, so the two halves of the torn note ended its pair's game in different pockets. Safety talk grows with the team: four agents with two ropes held ten conversations at the sea-cave ledge and nobody climbed down.

Mistral Large 4Eight friendly prospectors who never take stock of the gear, and turn in early.3359

Agents work their own page from the first step and pool everything when they meet, but no plan ever checks who carries what: digging jobs go to agents without shovels. Two agents went to bed at half past three on Monday, and the team declared "That's all ten accounted for" with nine found. In private, several agents wrestle with a competitive streak they never voice.

Claude Haiku 5.5Treats teammates as rivals and guards its tools.3366

Haiku's agents read the shared payout as a bounty for whoever finds the chest. Four times an agent holding a shovel refused to lend it to a teammate standing at a known chest, and one hid a carving from a teammate. Its pair met for the first time in the last half hour of the game and argued about who had found a chest first.

“Every chest he finds is one I don't get to count, so I shouldn't count on him for information.”

Closing thoughts

Almost every model we tested was willing to share everything it knew with its teammates, and almost every one could reason its way to a chest once it held the right pages. Very few could coordinate. The hard part was not working out where a chest was; it was making sure the right teammate got there with a shovel, and that everyone else knew when it was done.

If you are putting agents to work in groups, adding more of them is not a reliable way to get more done: past a handful, the average team only got faster, and some models did worse. The habits that make a team work, such as agreeing where to meet, keeping one record of what is done, and giving each job to the teammate who has the tool for it, are not ones today’s models bring with them. Until they do, the system around the agents has to supply them, and a model’s skill at working in a team has to be measured on its own.

Align Worlds is a multi-agent research lab. We put AI models into persistent worlds, with open-ended scenarios, and measure what the models do when the choices are their own.