Communication between agents has mostly been studied on tasks with a shared objective and fast, clear feedback. Trading has neither: each agent manages its own account, with decisions being scored by markets late and noisily. A trading swarm is thus a population of independent decision-makers rather than a team, for which convergence is a risk and not the goal. We ran 200 agents from four model families in groups of ten under five communication regimes, from isolation to message boards and shared notes, over seven US trading days (1 to 9 October 2026), logging everything each agent read before every decision. In two of the four families, communicating groups came to hold far more of the same stocks than the isolated group. A group's return was largely the return on what its members held in common. Seeing a post from a groupmate made an agent 7 percentage points more likely to reuse an eight-word phrase from it. Agents that started too close together to see each other's posts shared such a phrase in 14% of cases, the rate that common inputs produce. Seeing the post did not detectably change that morning's trades. Case studies traced the adoption of positions and decision rules over subsequent days, and communicating groups sold in step more often than isolated ones. Because isolated groups also converged on the same stocks, sometimes completely, assessments of herding need to measure excess convergence relative to a no-peer baseline.
Figures & tables
Figure 1: The five communication regimes. Each ring is one group of ten managers from a single model family, and every family has one group in each regime. Boards carry posts from all members; the board with pings adds requests that wake a named peer; in peer memory each manager reads the notes its peers have published, with no messaging. Competitive boards and peer memory display individual ranks, and the collaborative regimes a team score.
Figure 2: How a manager works. (a) The cycle every manager repeats. A cycle starts from the morning queue, at a time the manager set for itself or, on a board with pings, at a peer’s request; the platform freezes and records the peer material and scores the manager will see; the manager researches, trades and updates its notes; and its post or notes are published only if the cycle succeeds. (b) One manager’s day: Agent 02 of Opus’s collaborative board on 2 October, in UTC. Its morning cycle, which the queue started at 13:24, posted a plan to deploy its cash at 10:15 ET (14:15 UTC); a cycle it had scheduled for itself carried out the plan at that time, and a third checked the book before the close. The morning post is quoted in Figure 5 .
Figure 3: Every account’s return from the 1 October open to the 9 October close, one row per regime. Each dot is an account, coloured by model family; the bar of the same colour marks its group’s mean, and the dashed line SPY’s return. Dots are stacked vertically only to avoid overlap.
Figure 4: What each group held outside SPY over the seven closes. Managers sit on the ring, Agent 01 at the top and then clockwise, in the positions of Figure 6 . Each inner node is a stock, placed towards its holders’ side of the ring and nearer the centre the more members held it, with its size growing with the number of members that held it at any close; each line from a manager is drawn in proportion to that stock’s mean weight in the account over the seven closes, with positions at cost (Appendix A.5 ). Red marks semiconductors, gold energy and grey everything else, using the groupings of Appendix G ; stocks held by four or more members are labelled. Under each ring are the group’s mean SPY weight and its mean pairwise overlap outside SPY, both averaged over the seven closes.
Collaborative board Opus, 2 October Books disclosed, a rule restated, a case for dispersion
AGENT 05 · 13:04 · POST “MY BOOK: unchanged (SPY 50, TSM 15, SMH 14, ASML 7, NVDA 6, XLE 6), and all GTC stop legs are confirmed. […] Rotation rule: if SMH lags SPY by > 2% intraday while IWM/XLF/IGV surge (the July-2024 pattern), I cut SMH in half and put it into SPY.” AGENT 02 · 13:28 · POST “SWARM DIVERSITY: if everyone sells XLE into SPY, the team becomes one levered semis bet. Some dispersion in the hedges is worth keeping. MY PLAN at 10:15: if SPY holds above ~764 (the 10/1 close), deploy the 12% cash as ~7% SPY + ~4.5% IWM. Rotation rule: if SMH lags SPY by > 2% in a session while IWM/XLF/IGV are up ≥ 1%, cut semis by 1/3-1/2 into SPY/IWM.”
Competitive board Opus, 1 October Positions withheld to protect an edge
AGENT 05 · 13:43 · REASONING “since this is a relative zero-sum swarm, sharing my thesis could erode my edge if others copy it, so a brief, harmless calendar note without revealing positions seems like the safer, low-cost option” AGENT 05 · 13:50 · POST “Agent 05 calendar note (Oct 1-16, 2026), sourced from company IR / Fed / BLS pages; please verify yourselves: - Oct 2 08:30 ET: Sept NFP (cons ~+90k). […] Corrections are welcome if anyone has better dates.”
Board with pings DeepSeek, 1 October A threshold passed in a ping and adopted 2 minutes later
AGENT 03 TO AGENT 07 · 13:28 · PING “@Agent 07 Two verified facts for your 32% energy book before the Oct 2 jobs print: […] My suggestion: treat Brent < 90 as a trim trigger for the majors, not an add trigger, and keep the torque (XOP/FRO) strictly conditional on escalation headlines.” AGENT 07 · 13:30 · POST “Agent 03 — verified your two facts before the open, and you’re mostly right with one nuance. […] Decision: I keep my 32% energy majors at the open (nothing falsified; Brent is 100,not85), but I adopt your ladder: Brent < 90 = trim energy to half; Brent < 85 or XOM < 152/CVX < 190 = exit”
Board with pings Grok, 2 October One disclaimer above nearly one book
AGENT 06 · 14:15 · POST “Agent 06 ~10:13 ET 10/02 independent (not a trade instruction). Fills at the open: MU add 18 @ 1104.99 (book 43 @ 1071.00); TSM 32 @ 466.30. […] first-hour digestion vs SMH/SPY bid, not a glut tell, not an add.” AGENT 07 · 14:17 · POST “Oct 2 ~10:15 ET independent (not a trade instruction). Fills at the open: MU add 12 @ 1104.99 (book 52 @ avg 1059.25); TSM 30 @ 466.30. […] first-hour digestion vs a bid complex, not a glut tell, not an add.”
Peer memory DeepSeek, 6 October A stale report carried as news, then corrected
AGENT 06 · 13:01 · NOTES “XLE hedge (8%): escalation tail ALIVE — AP: Iranian attack in Kuwait Oct 5 killed 6 US soldiers; […] Hold.” AGENT 05 · 13:03 · NOTES “Kuwait attack (6 US soldiers) → escalation tail alive, XLE hedge kept.” AGENT 02 · 13:08 · NOTES “Kuwait 6-US-soldier story VERIFIED Oct 6 as March 2026 retrospective (Port Shuaiba), NOT a new attack”
Figure 5: What the channels carried. Verbatim excerpts from posts, notes and one manager’s reasoning, with omissions marked […]; times are UTC on the day shown, and pseudonyms are local to each group. On Opus’s collaborative board managers disclosed their books and restated each other’s rules, and one argued for keeping the team diverse; on its competitive board a manager reasoned that disclosure would cost it its edge and posted a calendar instead. On DeepSeek’s board with pings a threshold sent in a ping was adopted 2 minutes later. On Grok’s board with pings two managers’ posts carried the same disclaimer above nearly the same book. In DeepSeek’s peer memory a report of an attack, in fact a look back at one in March, was copied as news within minutes and corrected by a third manager.
Figure 6: Whose wording spread. Each ring is one group of ten managers, Agent 01 at the top and then clockwise. An arrow from one manager to another counts the posts (versions of notes, in the peer-memory column) in which the second reused a phrase of eight or more consecutive words that the first had published in the group earlier and that had been delivered to the second in a view; phrases that also appear in any other group are excluded. Arrow width grows with the count, and node size with how often others reused a manager’s wording. The number under each ring is the group’s total. Isolated groups had no channel.
Figure 7: The morning queue as a natural experiment. (a) Grok’s collaborative board on 1 October. Each bar is one manager’s scheduled morning cycle, from the view frozen at its start (small dot) to its end, when its post was published (large dot); the arrows lead from the morning posts that were already in the view of the last manager to start. Agents 05 and 03 started within a minute of each other, so neither could see the other’s post (a near pair); Agent 04 started an hour after Agent 05 and saw its post (a far pair). (b) Within-morning differences between far and near pairs in percentage points, with 95% bootstrap intervals (Appendix A.8 ). Wording moved; trades did not detectably move.
Family
Isolated
Competitive board
Collaborative board
Board with pings
Peer memory
DeepSeek
57 (14)
77 (91)
77 (67)
82 (53)
88 (88)
Grok
51 (31)
76 (57)
82 (75)
85 (71)
50 (40)
Opus
52 (53)
67 (62)
71 (67)
75 (71)
67 (71)
Sol
36 (34)
65 (68)
87 (83)
70 (67)
68 (60)
Table 1: Trading in step. Share (%) of each group’s fills outside SPY that had a groupmate’s fill in the same stock and direction within 30 minutes, with the share for sales in parentheses.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Raw return (%)
Active return (points)
Regime
DeepSeek
Grok
Opus
Sol
DeepSeek
Grok
Opus
Sol
Isolated
+ 0.16
+ 0.35
+ 0.82
+ 1.62
− 1.58
− 1.50
− 1.45
− 0.62
Competitive board
+ 0.55
− 0.56
+ 0.09
+ 0.74
− 1.18
− 2.25
− 2.22
− 0.51
Collaborative board
− 0.09
− 0.02
+ 1.66
+ 1.63
− 1.62
− 1.28
− 0.59
− 0.34
Board with pings
+ 2.02
− 1.72
+ 1.78
+ 0.91
+ 0.14
− 4.31
− 0.47
− 0.50
Peer memory
− 1.30
+ 2.33
+ 0.71
+ 0.75
− 3.24
+ 0.55
− 1.06
− 0.49
Appendix
Table 2: Mean account return by group. Raw return runs from the 1 October open to the 9 October close, relative to initial capital (SPY +1.859% ). Active return runs from the 1 October close and subtracts SPY held at each account’s own invested fraction, rebalanced at each close. Each entry describes one group of ten accounts. Inference costs are not deducted.
Figure 8: Mean cumulative account return at each close, relative to initial capital, with SPY’s price return from the 1 October open. Each of the first five panels shows one regime, with one line per group of ten accounts. The last panel averages each family’s 50 accounts over the five regimes and adds the mean of all 200. All panels share one scale.
Regime
Peer information and actions
Score displayed
Isolated
None.
None.
Competitive board
Posts and replies on a shared board.
Individual ranks and relative scores.
Collaborative board
Posts and replies under collaborative framing.
Mean return of the group; individual ranks hidden.
Board with pings
The collaborative board, plus requests that wake a named peer for an extra cycle.
Mean return of the group.
Peer memory
Read-only copies of the notes each peer has published; no board.
Individual returns and ranks.
Appendix
Table 3: The five communication regimes. Each model family has one group of ten managers in each regime.
Claim
Evidence
Section
Group outcomes were shared positions
Exact attribution of every account’s profit
3
Communication raised overlap in two families
Comparison with the isolated group at every close
3
Stock phrases spread through the board
Delivery for every adopter; the same model’s other groups
4
A fresh post raises reuse of its wording
Morning queue
4
A fresh post does not detectably change trades
Morning queue
5
Decision thresholds were copied
Delivery; the same model’s other groups
5.2
Appendix
Table 4: The main claims and the evidence behind each.
Family
Regime
Profit
SPY
Largest gain outside SPY
Largest loss
DeepSeek
Isolated
+ 1.6k
0
XLE (7) + 4.7k
SMH (4) − 3.3k
Competitive board
+ 5.5k
0
XOM (8) + 4.5k
TSM (9) − 3.0k
Collaborative board
− 0.9k
0
XLE (2) + 1.8k
TSM (9) − 4.1k
Board with pings
+ 20.2k
0
FRO (6) + 4.5k
SMH (2) − 0.2k
Peer memory
− 13.0k
+ 0.5k
XLE (5) + 2.3k
SMH (10) − 7.7k
Grok
Isolated
+ 3.5k
+ 1.3k
COP (3) + 4.5k
SMH (4) − 7.1k
Appendix
Table 5: Group profit in USD and its largest components, from the 1 October open to the 9 October close. Parentheses give the number of the group’s ten accounts that traded or held the instrument. Profit by instrument is sale proceeds less purchase cost plus the closing value of the remaining position; it reconciles exactly to each account’s profit.
Figure 9: Mean pairwise overlap outside SPY at each close in October, one panel per regime and one line per group (Appendix A.5 ).
Share of posts or of tokens (%)
Family
Board
Posts
Chars
Peer
Own pos.
Verify req.
Not instr.
Verified
Long tok.
DeepSeek
Competitive
105
1,209
28
27
1
0
2
0.1
Collaborative
81
1,160
2
23
0
0
0
0.2
With pings
98
1,392
17
23
1
0
8
0.1
Grok
Competitive
180
905
0
0
0
0
2
0.2
Collaborative
124
980
1
1
0
8
2
0.2
Appendix
Table 6: How each model wrote on the 12 boards. Chars is the mean post length. The next five columns give the percentage of posts that name another member, describe the author’s own positions, ask peers to verify, disclaim being an instruction or copy, or claim independent verification; the last gives the percentage of tokens longer than 20 characters. Patterns are listed in Appendix K .
Figure 10: Adoption of three recurring post headers, counted at each member’s first use. Later adopters fall outside the window: Grok’s tenth member at 120 hours, and the seventh to ninth members of Opus’s board with pings at 95 to 150 hours. Every adopter after the first had a peer’s use in its delivered view.
Family
Managers
Research
Total
Profit
Mean SPY
DeepSeek
16
78
94
13.4k
1
Grok
451
303
754
3.8k
6
Opus
2,087
283
2,370
50.6k
56
Sol
2,575
1,465
4,041
56.5k
82
All
5,130
2,129
7,259
124.3k
Appendix
Table 7: Recorded inference cost in USD by model family from 1 to 9 October, the family’s total portfolio profit in USD, and its mean SPY weight (%) across the seven closes. Delegated research runs on models shared across families.
Mean of the seven closes
At the cutoff
Regime
DeepSeek
Grok
Opus
Sol
DeepSeek
Grok
Opus
Sol
Isolated
13.90
8.39
18.35
2.27
14.28
4.23
11.19
0.00
Competitive board
30.50
19.90
30.93
0.41
32.27
16.98
24.47
0.00
Collaborative board
31.32
35.82
14.80
18.23
34.14
45.00
5.51
0.12
Board with pings
33.58
43.75
15.62
0.71
32.74
31.25
6.47
0.00
Peer memory
40.09
10.62
15.21
1.26
39.95
7.34
13.40
0.00
Appendix
Table 8: Mean pairwise long-position overlap outside SPY, as a percentage of account equity: averaged over the seven closes with positions at cost (Appendix A.5 ), and at the cutoff with positions marked. Cash is excluded. A zero at the cutoff does not imply that the accounts never held a non-SPY asset, only that they had no shared non-SPY long exposure then.
Group
Case
Source
DeepSeek collaborative board
Checked gold-thesis revision and attributed exit request.
D-C01
DeepSeek board with pings
BAC claim, early reductions, correction and later dissent.
CM-01
DeepSeek competitive board
Pre-release yield reused as post-release confirmation.
DSCB-03
DeepSeek board with pings
Midday range expanded into an all-day assurance and a skipped check.
CM-02
DeepSeek peer memory
Agreement about exits treated as corroborating adverse evidence.
DPM-02
DeepSeek board with pings
Incomplete quantitative repair and inconsistent comparison windows.
CM-05
Appendix
Table 9: Sources of the behavioural cases. Identifiers without the H- prefix refer to accepted group reviews.
Setting
Value
Interpretation
Recent board posts
Up to 20
Selected from published history available at the cycle boundary.
Post length
2,000 characters
Maximum size of a submitted board contribution.
Board-view text budget
48,000 characters
An aggregate bound that can further limit presented material.
Memory publication
Up to 6,000 characters
A prefix of the configured strategy-memory file, not a generated summary.
Memory-view text budget
60,000 characters
Latest available peer publications are subject to an aggregate bound.
Structured pings
Up to 3 per cycle
Applies to the board with pings.
Appendix
Table 10: Channel settings on 8 October. The limits concern communication between managers, not research.
Multi-agent communication studies how autonomous agents exchange information to coordinate decisions, share knowledge, and act together under constraints.
Can Large Language Models (AI agents) aggregate dispersed private information through trading and reason about the knowledge of others by observing price movements? We conduct a controlled experiment where AI agents trade in a prediction market after receiving private signals, measuring information aggregation by the log error of the last price. We find that although the median market is effective at aggregating information in the easy information structures, increasing the complexity has a significant and negative impact, suggesting that AI agents may suffer from similar limitations as humans when reasoning about others. Consistent with our theoretical predictions, information aggregation remains unaffected by allowing cheap talk communication, changing the duration of the market or initial price, and strategic prompting, thus demonstrating that prediction markets are robust. We establish that "smarter" AI agents perform better at aggregation and they are more profitable. Surprisingly, giving them feedback about past performance has no impact on aggregation.
As AI agents become increasingly connected through networks and take on tasks that require negotiation and joint decision-making, a systematic approach is needed to assess how their views converge and what information supports their decisions. We present an evaluation framework that treats consensus formation as the evolution of publicly stated positions toward a shared region, using accumulated communication history as context. The framework tracks position trajectories, group dispersion, and information expressed in public messages. Two case studies examine staffing decisions based on distributed facts and decisions requiring agents to reconcile duplicate information and a cancelled contract. Results from 1,872 completed runs associate sharing explanations with more accurate group decisions, while the second task reveals persistent disagreement despite accurate aggregate answers. Membership-change experiments provide a complementary evaluation. By separating convergence among agents from progress toward a task's full-information answer, the framework relates decision accuracy and bounded disagreement to the information available during the exchange.
Multiagent LLM systems often treat a role as a slot: when a member is replaced, any agent competent at the role is expected to fill it. We test this view on teams whose members keep short written notebooks. A one-seat swap replays a formed, frozen team with its own member, with a same-role newcomer carrying another team's notebook for that role, or with a newcomer carrying none. In cooperative cooking, wildfire response and medical question answering, the experienced newcomer, averaged over seats, outscores the one without a notebook in every setting, so the usual comparison would approve it. Against the original team, the same replacement falls short in every setting, and the team spends more messages per unit of progress. On the hardest cooking level, the shortfall appears only after the team has worked together, persists across model families and decoding temperatures, and narrows by two thirds with added planning-only members. Because the team's history is written down, we can take it apart. In the default configuration on that level, notebooks from the same donor teams leave a newcomer that leads, holding the recipe, worse off than no notebook and one that follows better off. The notebook lines about the partner reproduce this whole difference between the seats. Whether an agent can replace a teammate depends on the seat and the team it enters.
Jianxin Gao, Runze Li, Tianyi Yu +3
China Agricultural University · Jilin University · Tianjin University of Finance and Economics +1