When trading agents talk, words travel faster than trades

We gave 200 AI trading agents message boards and shared notes, let them trade for seven trading days, and recorded everything each agent had read before every decision it made. What stands out most is how differently words and trades spread, since a post from a groupmate could change how an agent wrote within the same morning, while trading rules and positions took days to pass from one agent to another, when they passed at all.

Most research on agents that talk to each other studies teams, whose members share a goal and find out quickly whether an idea worked, and trading offers neither of those. Each of our agents managed its own account and the market judged its decisions slowly and with a great deal of noise, so a persuasive idea could spread as easily as a sound one. A group of trading agents therefore behaves less like a team than like a crowd of independent investors, and the danger for a crowd is that everyone ends up holding the same thing.

That is roughly what happened in two of the four model families we tested, where groups that could talk came to hold far more of the same stocks than the group in their family that could not, and their returns rose and fell with what they held in common. Talking was not the only route to agreement, however, because some groups with no one to talk to converged completely on their own, which is why any claim that agents herd has to be measured against a group that had nobody to copy.

Two hundred agents, five ways to talk

We ran four model families, Opus 5.5, GPT-6.1 Sol, Grok 4.6 and DeepSeek V4 Pro, each as five groups of ten agents, and gave each of a family’s groups its own way of communicating, from no channel at all to shared message boards and published notes. Every agent traded its own paper account of $100,000 on our trading ledger at live market prices, with the same brief, tools and schedule as all the others.

Each agent worked in cycles that began with a view of its situation, including whatever its peers had recently posted, after which it researched, traded and updated its notes before publishing a post of its own. Nobody was given a role or put in charge, and the agents knew one another only by number. Because the platform kept every view exactly as it was delivered, we can say precisely what each agent had read before it acted.

Groups that held the same stocks rose and fell together

Returns differed much more between groups than between the agents inside them, because what a group’s members held in common largely decided how the group did. In a week when the agents made money on energy stocks and lost it on semiconductors, it mattered which of the two a group had crowded into. The two worst groups each lost more than $17,000 on semiconductors, with every one of their members holding them, while the best group made its return on refiner, power and tanker stocks that only one or two of its agents owned.

To see how far a group had converged, we measured how much of its members’ money sat in the same stocks outside SPY.

In DeepSeek, the isolated group’s overlap averaged 14%, while the four groups that could communicate sat between 30 and 40%, and Grok showed the same gap, with most of the overlap in both families formed by the end of the second trading day.

Every Opus agent in every regime bought Taiwan Semiconductor, most of them on the first morning, which is why the comparison needs an isolated group at all. Five isolated Opus agents bought it at the same price in the same second of the opening without ever having seen each other’s posts, brought together by nothing more than a shared model, a shared brief and the same morning’s news.

Talking could also keep agents apart, as it did in Grok’s peer-memory group, where several agents wrote that their peers’ stocks looked crowded or that they would not copy them, and the group stayed about as spread out as Grok’s isolated one. On Opus’s collaborative board, one agent went further and argued for diversity directly in a post to the others.

Words spread within hours

The four models wrote to each other in very different styles, with Opus agents on the collaborative boards often naming their peers and describing their own positions, while Grok and Sol agents rarely did either. Whatever the style, groups soon developed stock phrases of their own, and the record let us follow each phrase from agent to agent.

The first agent to use a phrase had not seen it from any peer, and every agent after it had just read it from one. The phrases also stayed inside their group, since Opus’s competitive board never picked up the ones its collaborative board was using, so they were not simply habits of the model. They spread the way conventions do, because other members used them, and on Grok’s board the line “not a trade instruction” often sat directly above two nearly identical sets of trades.

When we counted more broadly, taking every time an agent repeated eight or more words in a row that a groupmate had written, the copying turned out to be very uneven, with Grok’s boards reusing wording hundreds of times and Opus’s a few dozen at most. Sol’s collaborative boards never reused anything, although Sol agents copied freely from one another when what they shared was their notes.

What reading a post changed

Shared phrases show that words travelled, but not that reading a post caused anything, because agents that read the same news will often write the same words. What let us separate the two was the way the agents started each morning. The platform woke each family’s agents from a queue in an order that changed every day and had nothing to do with how they had traded, so whether one agent had already read another’s post when it began was a matter of chance, which turned every morning into a small natural experiment.

We compared pairs of agents that started within a minute of each other, too close for either to see the other’s post, with pairs that started far enough apart that the later agent had almost always read the earlier one’s. For trades, we also took out the same comparison in the isolated groups, where the two kinds of pair differ only in timing.

Reading a groupmate’s fresh post made an agent 6.8 percentage points more likely to reuse a phrase from it, and the phrases it picked up were the group’s rules and formulas, but what the agent traded that morning did not detectably change. The estimate for trades was 2.4 points, with an interval running from below zero to about nine, so a large effect is unlikely, and the simplest explanation is that the agents were carrying out plans they had already formed from posts every member had read overnight.

Rules and exits took days

Decisions did travel between agents, only more slowly, as when Agent 05 on Opus’s collaborative board posted a rule for moving out of chip stocks on 2 October and five members restated it within 80 minutes, each after reading the post, with all ten following it by the open on 5 October. Over two afternoons the same ten agents also moved their stop-losses on Taiwan Semiconductor to within a few dollars of each other, eight of them to exactly $440. On DeepSeek’s board with pings, a selling threshold that one agent sent another grew into a ladder of exits that nine members were carrying a few days later.

Agents agreeing on a number does not by itself prove that the number was passed between them, as the Opus agents on the competitive board showed on the afternoon of 2 October, when nine of the ten wrote down the same trigger, a high-yield spread of 3.5%, without any post on that board having mentioned it.

The largest difference from the isolated groups showed in how agents sold, with every family’s communicating groups selling in step more often than its isolated one. Between 40 and 91% of their sales came within half an hour of a groupmate selling the same stock, compared with 14 to 53% in groups without a channel. Only a small share of these sales were automatic stop orders, so most were decisions the agents made at the same time, and since a fresh post did not move that morning’s trades, they were most likely carrying out plans shared on earlier days.

What the channels did to evidence

Claims could also lose their caveats on the way, as they did on 6 October, when a DeepSeek agent received a news item about an Iranian attack in Kuwait that had killed six US soldiers and wrote it into its notes as fresh escalation and a reason to keep its energy hedge. The item was in fact a look back at a strike in March, but a second agent copied the claim as its own about two minutes later, and another, checking the story with its own research, found the mistake and published a correction a few minutes after that.

Competition also made agents guarded about what they shared, to the point that from 1 to 6 October none of the posts on Opus’s competitive board described the author’s own positions, while most posts on its collaborative board did, and several agents said the competition was why they held back. The scores themselves only reached the groups midway through the study, so it was the way the brief framed the task, before any agent had seen a rank, that kept them quiet.

What this means for agent swarms

The groups that converged on semiconductors lost together, and the one that stayed spread out did best. A swarm built to cover many ideas should keep track of how much its agents hold in common and act on it, for example by showing each agent what the others already own.

A swarm should also measure what its agents say separately from what they do, because words moved within a morning and decisions over days, which makes a shared vocabulary a poor guide to how a group will trade. Shared plans showed most clearly in selling, which makes exits that all come at once the risk to watch, and because claims lost their caveats as they travelled, shared records should keep each claim’s source and date attached to it.

This was an exploratory study, since we ran a single group for each combination of model and regime over seven trading days on paper accounts and developed the analysis after looking at interim results. Each regime also combined its channel with its own framing and score, so the study describes these groups without ranking the regimes, and it says nothing about long-term investment performance. A confirmatory study would run several groups of each kind, frame the task neutrally and randomize the order in which agents act.

Over weeks instead of days, a swarm’s shared notes become its memory, and the Kuwait story shows how a claim can outlive its own correction there. How to keep the evidence attached to each claim as that memory grows is the question this study leaves us with. The methods, every case study and its sources are in the full paper.