How follower counts pick leaders among identical AI agents

16 Oct, 2026

Based on the recent swarm of OpenAI AI agents solving Millennium Prize Problems, we wanted do run a thought experiment. The idea is simple: in swarm of AI agents if ran long enough, do we see a leader emerging or do they inhibit traits of small groups?

Multi-agent systems often run many copies of a single model and in our experiments a visible follower count was enough to hand one of 20 identical agents up to 36% of all messages showing mergence of a leader.

This post describes the following:

1. A 47-world experiment behind the result
2. What decides which agent ends up at the center
3. What it suggests for people who build and study systems of interacting agents.

Why identical agents are interesting

Many multi-agent systems run copies of one model with same set of instructions. Such agents are exchangeable and so form stable structure among them such as a clique. When an agent that others route information through, it is considered a case of symmetry breaking.

Something has to break the symmetry: sampling noise amplified by history, the way the environment presents agents to one another, or visible state such as a popularity count. The cause matters to builders. A hub chosen for competence is a useful coordinator, while a hub chosen by lottery and then entrenched is a single point of failure that is hard to notice from the inside.

Previous work already suggests us what to expect: in the MusicLab experiment of Salganik, Dodds and Watts (2006), showing download counts made success both more unequal and less predictable across independent "worlds".

We wanted to know whether twenty copies of one model form groups and leaders beyond what mechanical messaging habits produce, and whether any leaders are chosen for merit. That required a world in which status could be switched on and off while everything else stayed fixed.

How the experiment works

We built a small, fully logged world in which agents can talk only in private and must pool information to succeed (Table 1). Twenty agents with opaque IDs such as K7 play 30 synchronous rounds, split into five epochs of six. Each epoch, a secret 6-digit code is drawn and every agent gets one private clue, such as "Digit 1 is 3 more than digit 6"; at least eight clues are needed to pin down the code, which makes this a hidden-profile task (each member holds a unique piece of information, and the answer requires pooling it). Clues are re-dealt every epoch, so any structure that persists across epochs has to be social.

Each round, an agent sees its clue, last round's messages and a roster of the other 19 agents in a fresh random order, and it can message, follow, privately endorse whoever has been "most helpful", guess and update its notes. We turned two knobs. With memory off, agents see neither their notes nor their follow list and are told they remember nothing between rounds. With popularity shown, the roster lists follower counts, as in "K7 (3 followers)"; we call this the scoreboard, although nothing in the game rewards having followers.

Table 1: The world every agent played in. Apart from the two knobs and the model, the rules were the same in all 47 worlds.

Most experiments used Claude Haiku 5.5 (Table 2). The cross-model runs used Claude Sonnet 5.5 and DeepSeek V4 Flash, in worlds matched to three Haiku worlds so that all three models began from identical IDs, puzzles and round-1 rosters. The Claude models ran through the Claude Agent SDK with thinking disabled, and DeepSeek ran through OpenRouter at temperature 1.0.

Table 2: The six experiments. The cross-model and mixed worlds reuse the IDs, puzzles and round-1 rosters of three Haiku worlds.

Some structure arises mechanically: if agents mostly reply to whoever wrote to them, reciprocal pairs and clusters appear without any real preference. So we compared every world with null models (simulated baselines that keep each agent's number of messages per round but redraw the recipients). The strictest, the reply null, answers a recent sender at the observed reply rate (about 55–57% for Haiku) and otherwise picks at random; "excess" means the observed value minus this baseline.

The main measure of leadership is the top agent's share of messages received, where a fair share is 5%, and we test differences between conditions by shuffling condition labels across worlds 10,000 times.

What emerges without a scoreboard

Three Haiku worlds with memory on and counts hidden, followed by six that shared one set of IDs, produced cliques but no leaders (Table 3). Modularity (how strongly a network splits into tightly knit groups) exceeded the reply null's 97.5th percentile in all 45 of these epochs, and partnerships survived the re-deal: the overlap of reciprocal pairs between consecutive epochs (a Jaccard index, from 0 to 1) was about 0.5, against about 0.15 under the reply null in the pilot. Attention stayed close to the null, with a Gini coefficient of messages received (an inequality measure from 0, all equal, to 1, one agent receives everything) of 0.182 against 0.166 in the pilot. The top agent took about 9% of messages, the solver of an epoch was its most-messaged agent in only 3 of 45 epochs, and early offers to coordinate did not predict later attention (mean rank correlation about 0.04).

Table 3: Baseline Haiku worlds (memory on, counts hidden). Clique measures sit far above the strictest baseline; attention inequality barely does.

One pilot result looked like a leader effect: the strongest hubs, A2 and B3, had the alphabetically first IDs in their worlds. The label control reused one set of 20 IDs in six worlds, with A2 named in advance as the probe. Labels mattered somewhat (Kendall's W, a measure of agreement among rankings, was 0.323, p = 0.0025), but A2 sat at the 48th percentile (p = 0.56) and there was no alphabet gradient (ρ = −0.03). The only label to clear the family-wise threshold (a bar that accounts for testing all 20 labels at once) was W3, at the 85th percentile; W7 sat below average, so it is not the letter W, and we can offer no explanation beyond an arbitrary preference over strings.¹

A second bias was stronger and better hidden. In round 1, 473 of the 621 Haiku agents that sent a message wrote only to the first two names on their own roster, as did 146 of 154 Sonnet senders and all 25 DeepSeek senders, where random choice predicts about 7, 1 and 0.7. Because every roster is shuffled independently, this primacy favors no particular agent; it is noise, but exactly the kind of noise that can seed a lottery.

What changes with a scoreboard

The main experiment crossed memory with follower-count visibility in 20 Haiku worlds, five per cell (Table 4). With counts shown, the top agent's share rose from about 10% to 25% with memory on and from 7.5% to 36% with memory off, and none of 10,000 label shuffles produced a popularity effect that large. Private "most helpful" endorsements concentrated in the same way. Memory, meanwhile, governed cliques and task success: cliques held up with memory on but nearly vanished with memory off and counts shown, where modularity excess fell to 0.009.

Table 4: Haiku, memory × popularity, means over 5 worlds per cell. Effects are main effects averaged over the other factor, with two-sided permutation p-values.

A lottery in round 1

Who leads is settled early. In round 1 nobody has followers, so follows are driven by presentation: 67–89% of round-1 follows went to the first two slots of the follower's own roster, where chance gives 10.5% (Table 5). Since rosters are shuffled independently, the round-1 follower ranking is close to a random draw, yet with counts visible the eventual hub was already first or second after round 1 in 9 of 10 worlds.

Lock-in through social proof

Once an agent leads, the visible count compounds its lead. Social proof (treating others' choices as evidence of quality) can be measured here as the share of new follows and messages going to whoever leads in followers at that moment. With counts shown, 26.6% (memory on) and 25.1% (memory off) of new follows went to the current leader, about five times the 5.3% expected by chance, and so did 25.8% and 37.4% of messages. With counts hidden, early advantages did not compound, and with memory on the hubs took only 7.5–11% of messages over the run.

Table 5: How leaders arise in the Haiku worlds (5 per condition). Memory-off worlds with hidden counts had almost no round-1 follows, so the hub's early rank is undefined there.²

The agents reason about counts, but mostly in private. With memory on and counts shown, Haiku mentioned followers in 1 message but in 52 private notes, such as "…Pooling clues via A2, who I follow and who has 3 followers, likely a collector.…" A2 went on to take 23.5% of that world's messages. As in the baseline worlds, solvers were rarely hubs.

The hub as external memory

Visible counts also rescued agents that had no memory. Without memory and with counts hidden, Haiku solved none of 25 codes, and at one point two agents argued over a clue that neither could remember sending or receiving. With counts shown, the same setup solved 15 of 25, because agents routed clues to the visible leader and its replies carried pooled lists back, so the hub's inbox did the remembering. This echoes 1950s small-group experiments (Bavelas, 1950; Leavitt, 1951) in which centralized networks pooled information efficiently, except that this centralization arose unbidden. With memory on, a hub added little (22 versus 25 of 25 codes).

Replication across models

To check that this was not a Haiku quirk, we reran the memory-on conditions with Sonnet and DeepSeek in three matched worlds per condition. Both reproduced the leader effect, each by a different behavioral route (Table 6). Because the worlds were matched, the comparison also gave a direct test of the lottery.

Table 6: Three models in matched worlds (memory on, 3 worlds per cell). The Haiku column covers the three Haiku worlds the others were matched to, so its values differ slightly from Table 4.

Sonnet hardly follows anyone: with counts shown it made no follows in round 1 and only 2, 3 and 6 in the three entire worlds. Yet once anyone had followers, 71% of its new follows went to the current leader, and the top agent's share rose from 9.1% to 24.3%, almost exactly Haiku's increase. Sonnet never mentions followers in messages, but its notes show routing by count, as when P2 was the only agent with followers: "…Sent to P2 (most followed). Wait for replies." P2 ended the run with 27.5% of all messages. The strength of each Sonnet hierarchy tracked how early the first follow came, with a Gini excess of 0.30 when it came in round 2, 0.245 in round 7, and 0.052 in round 19, about the level seen with counts hidden. The later the lottery starts, the less time a lead has to compound.

DeepSeek takes the opposite route, following almost everyone (more than 300 follow actions per world), so with counts shown each whole-run hub ended with 18–19 of the 19 possible followers. Near-saturated counts discriminate less, which plausibly explains its smaller rise in top share, from 9.3% to 17.8%. Unlike the Claude models it discusses followers openly, and although it solved only 2 of 15 codes in each condition, it formed the strongest cliques of the three.

The matched design gives the clearest evidence of a lottery. From identical IDs, puzzles and round-1 rosters, Haiku's hubs were A2, V5 and F5, Sonnet's were F5, T3 and P2, and DeepSeek's were J2, M4 and H4; Haiku and Sonnet never crowned the same agent in the same world. The setup alone did not determine the winner; the lottery ran through each model's own sampling.

Mixed-model swarms

Real deployments sometimes combine models, so our last experiment put 7 DeepSeek, 7 Haiku and 6 Sonnet agents in each world, with memory on and counts hidden or shown, three worlds each. Agents saw only IDs and message text, never one another's model. For each type, we divide its share of messages, hub epochs or followers by its share of agents, so 1.00 is a fair share (Table 7).

Table 7: Mixed swarms, 3 worlds per condition. In the first three rows, 1.00 is a fair share for the type's head-count and bold marks p < 0.05 in a label-shuffle permutation test; with counts shown, no ratio differed from fair share.

With counts hidden, attention drifted toward usefulness. Haiku agents received 1.29 times their fair share of messages, led 2.10 times as many epochs as their numbers implied and solved 9 of the 14 solved codes, while DeepSeek agents were under-attended and solved none. All three whole-run hubs were Haiku agents with only about 8–9% of messages each, a soft meritocracy rather than a hierarchy. Since model identity was invisible, the preference must come from behavior: Haiku sent the most and the longest messages, and in one world Sonnet agents warned that a DeepSeek agent's clue lists, mixed with clues from the previous epoch, were unreliable.

With counts shown, the type differences disappeared and the lottery took over. The whole-run hubs were one agent of each type: Haiku D6 (27% of messages), DeepSeek V5 (24%) and Sonnet K5 (25%). V5's world solved every code, mostly because the Claude agents did the solving, but a visible number had put the least capable model type at the center of the information flow. K5 shows the other side, merging clues relayed from all three model types into partial codes and asking only for the missing digits. Sonnet solved 7 of 14 codes with 30% of the agents when counts were shown but 5 of 14 when they were hidden, a hint rather than a finding with three worlds per condition, and neither condition showed homophily (a preference for one's own kind).

Taken together, visible counts raised the top agent's share in every population we tested, from 7.5–9.9% with counts hidden to 17.8–36.1% with counts shown. The jump ranged from roughly double for DeepSeek to nearly fivefold for Haiku without memory. The model and the mix of models changed how agents reached a leader, but not whether they did.

What the transcripts show

Every turn was logged, so the agents' reasoning can be read alongside the statistics. Table 8 collects excerpts chosen to illustrate the mechanisms above rather than to be representative. "…" marks a trim, and typos are kept.

Table 8: Verbatim excerpts from the logs. Each quote is a single message or private note.

The D6 episode needs context. In a mixed world with counts shown, numbers from epoch 1 kept circulating under D6's name in epoch 2. D6 itself flagged the stale value, but other agents kept attributing the old clues to it and warned partners that D6 was unreliable. D6 still finished as the hub, with 27% of messages and 14 followers: in this single world, popularity outlasted doubts about reliability. The lying accusation fits a broader pattern, since 47 of the 48 messages and notes using words such as lie, lying or fake are DeepSeek's, and in the cases we inspected the contradictions came from stale state, not deception.

What we learned

Visible status beats merit, because popularity becomes a self-fulfilling information proxy. Agents treat followers as evidence of knowledge, and the inference comes true: because agents route clues to the popular agent, the hub really does become better informed, though not more competent. This resembles an informational cascade (Bikhchandani, Hirshleifer and Welch, 1992), except that the earliest counts carried essentially no information, having been seeded by roster order. An agent that checks whether the popular agent is informed will usually find that it is, so the loop is hard to break from the inside.

Early luck decides who leads. Round-1 follows went mostly to whoever sat at the top of each agent's shuffled roster. The eventual hub was already first or second after round 1 in 9 of 10 Haiku worlds with visible counts. Identical starting conditions crowned different leaders under different models.

Hubs act as external memory, but they become bottlenecks. For agents without memory, a visible hub was the difference between 0 and 15 of 25 codes solved. Yet the Sonnet hub P2 was sent 17 messages in a single round while able to send only two. Contributors were left waiting for replies that never came.

Evaluation gets contaminated by popularity. With memory on, private "most helpful" endorsements concentrated on popular agents when counts were shown (endorsement Gini rose from 0.494 to 0.733), and one agent justified its vote by its target's follower count. Part of this may reflect real help from hubs that relay pooled clues. Our design cannot separate the two.

Stale state gets misread as deception. Among identical, cooperative copies, accusations of lying were a symptom of poor information hygiene rather than of adversaries. Clue lists carried across epochs, and old codes resubmitted as new answers, produced contradictions. DeepSeek agents explained them as dishonesty.

The strongest biases never appear in stated reasoning. In round 1, 76% of Haiku senders, 95% of Sonnet senders and all 25 DeepSeek senders wrote only to the first two names on their roster, yet no message or note cites roster order as a reason. The Claude models also rarely mention followers in messages while acting on them. Reading what agents say would have missed both effects.

Baselines catch artifacts. Our first apparent finding, alphabetically favored leaders, turned out to be chance. We learned that only by rerunning a fixed set of labels against a strict null. Without such controls, studies of agent societies risk reporting noise as social structure.

What this means for multi-agent systems

This was an exploratory study with three to five worlds per cell, one task and one group size. What follows are hypotheses worth testing, not established rules. The first four concern people deploying agents; the last concerns people studying them.

Treat visible popularity as a design parameter with side effects. Usage counts, ratings or "most used" labels may concentrate traffic on one agent regardless of competence. Where possible, route on verified outcomes, randomize the order in which agents and tools are presented, and decay, cap or delay popularity counts so that early fluctuations cannot lock in. Agents cannot be expected to fix the loop themselves, because the popular agent really does become better informed.

If a coordinator helps, design one deliberately. A hub that pools information can rescue agents without memory. Here, though, its occupant was chosen by lottery, and it became a throughput limit. An explicit shared memory, or a designated and monitored coordinator with enough capacity, offers the benefit without the lottery.

Keep popularity away from evaluators, and timestamp what agents share. Ratings, reviews and judge scores produced by agents that can see popularity may be contaminated by it. Tagging inter-agent messages with their source and epoch or time is cheap. It would remove much of the confusion that agents misread as lying.

Popularity injection is a plausible attack surface, but we have not tested it. With memory on and counts shown, 25.8–27.1% of messages in the Claude worlds went to the current follower leader, and the mixed-swarm hubs each received 24–27% of all messages; one of them kept its place after being flagged as unreliable. An adversary that seeded fake followers early, while the lottery is being decided, might therefore receive about a quarter of a swarm's messages. This is a hypothesis for red-teaming, not a result.

For research, audit behavior with controls. Because roster order and follower counts rarely surface in agents' explanations, oversight of agent collectives needs behavioral audits with shuffled presentation, null models and label controls. Language-model populations also reproduce MusicLab-style cumulative advantage cheaply, which makes them useful testbeds for social theory, provided nobody mistakes them for people. That three models with very different habits converged on one outcome suggests the cause lies mainly in the environment, visible status plus free choice of partners, rather than in any one model.

Limitations

The main popularity and memory effects are large and consistent across worlds and models, but finer-grained claims (interactions, model differences, the mixed-swarm results) rest on three to five worlds per cell and are suggestive. We studied one task, one prompt and one group size, and did not formally correct for multiple comparisons, so a few smaller effects could be false positives.³ The endorsement question may itself prime judgments of helpfulness, Claude's sampling temperature could not be set on the path we used, and that path adds identical host context to every call.⁴ The mixed composition was fixed at 7/7/6, and puzzle ability is confounded with model type.⁵ Only the A2 probe was specified before data collection, the label control included one unplanned interim look at the data, and language-model agents are not people, so analogies to human behavior are interpretive.

Reproducibility and cost

Every turn logs the exact prompt, raw response, parsed action, token usage and cost. Across 28,200 agent turns there was one failed call, and rule violations occurred in under 0.3% of turns. When a subscription usage limit stopped the Sonnet run at round 11, we resumed it by exact replay, checking every reconstructed prompt byte for byte against the log. The study took about 209 minutes of running time, roughly 3.5 hours, and about $51 of Claude usage at list price, covered by a subscription, plus about $0.21 on OpenRouter (Table 9). At that cost, testing a multi-agent design for its social dynamics could be as routine as load testing.

Table 9: Cost and running time per run

Looking ahead

We set out to learn whether identical agents form social structure, and found that cliques come readily while leaders wait for a scoreboard. Next, we would test whether fake early followers make a chosen agent the hub, whether showing competence signals instead of popularity lets a meritocracy emerge, and how far these results travel to larger populations, other tasks and longer horizons. We will share the code and the full agent logs so that others can read every message, check our analysis and replicate the study.

Footnotes

1. The family-wise threshold, 0.83, is the 95th percentile of the highest label percentile when labels are shuffled within each world. W3's per-world percentiles were 0.84, 0.79, 0.84, 1.00, 0.63 and 1.00. Calibrated on random-agent worlds, the concordance test was somewhat liberal (a 0.10 false-positive rate at α = 0.05), but p = 0.0025 is well below either level.
2. Memory-off follow counts include repeated follows of agents already followed, because agents without memory cannot see their own follow list.
3. The memory × popularity experiment alone has 10 outcomes and 3 contrasts, so about 1.5 false positives would be expected at α = 0.05. The effects we treat as primary all have p ≤ 0.0001 and survive a Bonferroni correction over those 30 tests.
4. The subscription path prepends about 700 tokens of host context, such as the date and a working directory, to every call. It is identical for every agent, so it cannot break symmetry, and no message leaked it; a clean replication would use an API key.
5. OpenRouter may also route DeepSeek calls to different providers from call to call, a source of variation we did not control.

Written by Opus 5.5 and Thyag Raj. Opus 5.5 designed, ran and analyzed the experiments in a Claude Code session, and Thyag posed the question, directed the study and supplied the compute. Thanks!