Test-Time Communication in Trading Agent Swarms

• In two of four model families, communicating groups had greater portfolio overlap than isolated groups, and shared holdings accounted for much of each group’s return.Source: §1 · p. 2
• A groupmate’s fresh post raised reuse of its wording by 7 percentage points.Source: §1 · p. 2
• A fresh post did not detectably change that morning’s trades, while traced decision rules took one to four trading days to spread and communicating groups sold in step more often.Source: §1 · p. 2
• Competitive framing coincided with Opus managers withholding positions, while communicating groups in two families requested less research without establishing a division of work.Source: §1 · p. 2
• Every Opus manager bought TSM, with or without peers, showing that substantial convergence can occur without communication.Source: §1 · p. 2

2.1 Managers and cycles

• Each manager runs in a continuous context, receives a view of available peer material at cycle start, then researches, trades, updates notes, and publishes after a successful cycle.Source: §2.1 · pp. 2–3

2.2 Communication regimes

• Each communication regime combines a channel, its framing, and a score display, while managers retain authority over only their own accounts.Source: §2.2 · p. 3

2.3 Separating influence from common inputs

• Near pairs started too close together to see each other’s fresh posts, while far pairs usually could; the study subtracts the same timing comparison in isolated groups.The morning queue order changed daily and was unrelated to managers’ prior placement, returns, or activity.Source: §2.3 · p. 4
• Because each model–regime combination had one group, regime differences are confounded with group history, so the study describes realized groups rather than ranking regimes.Behavioural case studies came from chronological reviews, and authors checked each case against source records.Source: §2.3 · p. 4

Returns and attribution

• 0.622% combined equity growth lagged SPY’s 1.859%; 35 of 200 accounts beat SPY, and losses were concentrated in positions held by multiple group members.Sol’s 0.49-point shortfall against SPY was smaller than other families’ 1.16–1.76 points, reflecting its 82% average SPY allocation rather than better selection.Source: §3 · p. 4
• USD 17,418 and USD 19,748 in semiconductor losses came from every member of DeepSeek’s peer-memory and Grok’s board-with-pings groups, respectively.The two best groups lost USD 163 and USD 4,197 on semiconductors; group returns were tied to shared holdings, while Grok’s highest return included stocks held by only one or two managers.Source: §3 · p. 4

Holdings and overlap

• 14% average overlap in DeepSeek’s isolated group rose to 30–40% in its communicating groups; Grok’s 8% rose to 20–44%.Most overlap formed in the first two trading days; the figure maps holdings and mean pairwise overlap outside SPY across seven closes.Source: §3 · p. 4, Figure 4 · p. 6

Isolated convergence

• Every Opus manager bought TSM in all five regimes, including five isolated managers filling at the same price in the same second.This convergence in isolation shows why excess overlap must be assessed against a no-peer baseline.Source: §3 · p. 5

Framing and limits

• The design cannot distinguish communication framing from channel effects because each regime bundles both and has one group per family.Overlap varied by framing: Grok’s collaborative boards exceeded its competitive board, while Opus’s collaborative groups overlapped less than its isolated group.Source: §3 · p. 5

Alternative responses

• Grok’s peer-memory group’s 11% weekly overlap and 7% cutoff overlap stayed near its isolated group’s 8% and 4%, despite inconsistent attempts to avoid crowded holdings.Source: §3 · p. 5

Writing styles

• Opus named peers in 47–49% of collaborative-board posts and described its own positions in 31–54%, while Grok and Sol rarely did either.Sol also had 9–24% of tokens longer than 20 characters, versus at most 0.2% for the other models.Source: §4 · p. 7

Phrase adoption

• On Opus’s collaborative board, “verify before use” reached seven members in about two hours and all ten by the next morning.On Grok’s board with pings, “not a trade instruction” reached eight members within 71 minutes and eventually all ten; later adopters had seen the phrase in their cycle’s delivered view.Source: §4 · p. 7

Measuring reuse

• Grok’s boards reused wording 230–383 times each, while Opus’s boards reused it 10–35 times; Sol’s peer-memory group copied notes 178 times.Reuse counted phrases of at least eight consecutive words first published by another member and already delivered to the reuser, excluding phrases found in any other group.Source: §4 · p. 7

Queue-timing comparison

• 6.8 percentage points more often, later posters reused an eight-word phrase when queue timing let them see the earlier post than when both started together.The 95% interval was 1.7 to 11.8 across 68 group-mornings; pairs starting together still shared new phrases in 14% of cases, reflecting common research results or status templates.Source: §4 · p. 7

5.1 Same-morning effects

• The queue comparison found no detectable effect of a fresh post on that morning’s trades, while case studies traced adoption of positions and decision rules over longer periods.The comparison distinguishes overall trade similarity from the effect of seeing a fresh post; its estimated far-minus-near difference was 2.4 points (95% interval −4.1 to 9.3).Source: §5 · p. 7, §5.1 · p. 7

5.2 Decision rules

• Three decision thresholds spread beyond the no-peer baseline, each reaching nine or ten members within one to four trading days.Examples include an Opus rotation rule and TSM stops converging on USD 440, and a DeepSeek Brent trigger becoming a 90/85 ladder; the high-yield 3.5% trigger also appeared among isolated competitive managers without a matching post.Source: §5.2 · pp. 7, 10, Figure 5 · p. 8

5.3 Timing of trades and exits

• Communicating groups had more same-stock, same-direction fills within 30 minutes than isolated groups in 15 of 16 comparisons, and all 16 had a higher share of sales in step.For sales, the shares were 40 to 91% in communicating groups against 14 to 53% in isolated groups; most in-step sales were not automatic stop orders.Source: §5.3 · pp. 10–11, Table 1 · p. 10
• Common schedules, news, and prices also synchronized isolated trades, but communicating groups had a higher share of same-stock, same-direction fills in most comparisons.Five isolated Sol managers sold MU within 21 minutes, each citing its own test.Source: §5.3 · p. 10

6 Evidence, disclosure and effort

• Some claims became more certain as managers relayed them, while Opus managers withheld positions under competitive framing and communicating groups in two families requested less research without establishing a division of work.These patterns cover changes in qualification, position disclosure, and research effort.Source: §6 · p. 11

6.1 Qualification of claims

• A stale report of a March 2026 attack was copied as fresh news, corrected by a fourth manager, yet remained described as a new attack in the original author’s notes through 9 October.The correction reached the original author’s view the next day.Source: §6.1 · p. 11
• Claims also strengthened without new evidence, including a yield window broadened to a full day and an incorrect BAC price claim that prompted two reduction orders before correction.Similar qualification loss occurred within isolated managers’ records, while communication allowed strengthened versions to reach additional decision-makers.Source: §6.1 · p. 11

6.2 Disclosure under competition

• None of Opus’s 40 competitive-board posts described positions, compared with 43 of 63 collaborative-board posts; disclosure varied across model families and later increased mostly for exits already made.At least three managers cited competition as their reason, and the pattern was not consistent across families.Source: §6.2 · p. 11

6.3 Research volume and division of labour

• DeepSeek and Opus communicating groups made fewer research requests than their isolated counterparts, but most stock-naming requests concerned stocks already researched by a groupmate.Managers rarely used targeted requests, and no reviewed group established a lasting division of research.Source: §6.3 · p. 11

Related work

• Machine-behaviour research examines interacting AI systems, including information diffusion, shared conventions, and collective biases emerging from local interactions [7] [8].Test-time communication studies find debate’s benefits vary by task and architecture [9] [10], while workspace teams can match several times as many independent agents on some open-ended tasks [11].Source: §7 · pp. 11–12
• Social-influence research connects observed decisions with convergence and diversity, while language-agent studies find adoption depends on confidence and presentation and that structural interventions can control herding better than prompt instructions [26].The paper characterizes its untrusted-content warning as an instruction rather than a structural control, and frames peer-effect identification as the reflection problem [15].Source: §7 · p. 12
• Agent-network experiments report robust wording convergence without reliable stance shifts from distributed sources, and persuasion studies distinguish direct exposure, peer relays, and belief probes [29] [30].Transmission-chain experiments also find small biases compounding as text passes between models [28].Source: §7 · p. 12
• Prior financial-agent systems route contributions through simulated markets, central allocation, or manager–analyst hierarchies, unlike this study’s independently controlled accounts [4] [16] [17] [31].A five-structure trading-agent study found structure effects depended on stocks and isolated agents converged as far as communicating ones [25]; this paper adds live markets, four model families, exposure records, and a queue-based natural experiment.Source: §7 · p. 12
• Live-market evaluation and memory research motivate examining market exposure and retained information, while this study distinguishes conversation history, personal files, peer notes, and shared boards by their exposure rules [5] [19] [20] [32] [33].The cited benchmark attributes most agent returns to passive market and style exposure, while this study reports active returns against an exposure-matched benchmark.Source: §7 · p. 12

8.1 Implications for swarm design

• Communicating groups in two families held more of the same stocks than isolated groups, while Opus converged on TSM without peers; the authors suggest measuring common exposure when designing swarms for diverse coverage.Groups converging on semiconductors lost together, while a dispersed group had the highest return.Source: §8.1 · p. 13
• Wording reuse and trading occurred on different timescales, so shared wording is limited evidence about trading; synchronized sales were the clearest difference against the no-peer baseline.The authors identify correlated exits as a particular concern for collective portfolio risk.Source: §8.1 · p. 13
• Because claims sometimes lost qualifications in relay, shared records should retain each claim’s source and time window and distinguish peer actions from supporting evidence.Research remained overlapping and targeted requests were rarely used; explicit task allocation is proposed for testing as a way to establish division of work.Source: §8.1 · p. 13
• The study cannot establish which communication regimes suit each model family, despite differences in how families used the tools.Sol accounted for 56% of recorded inference cost while holding mostly SPY.Source: §8.1 · p. 13

8.2 Limitations

• Each model–regime combination had only one group, so the study describes groups rather than ranking regimes; each regime also bundled framing, channel, and score display.The queue comparison covers morning cycles and assumes the isolated timing effect would hold in communicating groups without a channel.Source: §8.2 · p. 13
• Seven trading days in one market environment support detailed accounts of disclosure and convention formation, but not claims about durable investment performance.The study used paper accounts, did not deduct inference costs from returns, and was exploratory; recorded reasoning is not a complete account of decision causes.Source: §8.2 · p. 13

9 Conclusion

• Shared holdings accounted for much of each group’s return, while communicating groups sold in step more often than isolated groups.Peer exposure and comparisons with isolated groups help distinguish peer influence from similarities arising without communication.Source: §9 · p. 14
• Communication also affected the evidence managers used: some claims lost qualifications as they were relayed, and a correction failed to change the original author’s notes.The authors identify preserving claims’ evidence and discussion in swarm memory as a question for future work.Source: §9 · p. 14

Data and methodological transparency; A.1 Population and records

• The archive includes source records, reconciliations, linked case reports, analysis scripts, and review traces; blind recomputations reproduced returns, attribution, and several behavioral estimates within rounding or resampling error.The records are not publicly available, and informational inputs are described without proprietary supplier identities.Source: Data and methodological transparency · p. 14

A.2 What every view said

• Each account’s capital and order authority remained isolated, while peer content was labeled untrusted and could not authorize trades or override its mandate.Managers ran in continuous context with automatic compaction; weekend cycles and 133 orders remaining open after the pause were excluded.Source: §A.2 · p. 17, §A.1 · p. 17

A.3 Returns, exposure and attribution

• Active return compares an account with SPY held at that account’s invested fraction; it combines stock selection, timing, and execution.The benchmark is rebalanced daily, and all positions were long.Source: §A.3 · p. 17

A.4 A reference for noise

• Regime means are compared with 20,000 shuffles of regime labels among each family’s 50 accounts, but accounts within a group are not independent regime replications.Even a difference exceeding the shuffled reference may reflect a group’s shared history, and common market movements still couple groups.Source: §A.4 · pp. 17–18

A.5 Portfolio overlap

• Pairwise overlap sums the minimum long-only asset weights, excludes cash, and is averaged over a group’s 45 account pairs.Reported overlap excludes SPY or includes it; Figure 9’s close-level positions use cost basis over closing equity.Source: §A.5 · p. 18

A.6 Posts, conventions and wording

• Wording reuse counts a first-time reuse of at least eight consecutive words only when an earlier peer’s phrase appeared in the delivered view and was absent from other groups.Convention adoption likewise requires checking whether a peer’s earlier use appeared in the adopter’s delivered view.Source: §A.6 · p. 18

A.7 Trading in step, decisions, ranks and research

• A same-stock, same-direction fill is in step when a groupmate filled within 30 minutes; protective stops are excluded from decisions because they arm exits in advance.Research requests count as stock-related when they name a traded ticker or mapped company name.Source: §A.7 · p. 18

A.8 The morning queue

• Morning comparisons use within-group-morning far and near pairs; isolated groups estimate timing differences without a communication channel.Pooling mornings or classifying exposure by whether a prior cycle ended can bias comparisons; alternative timing thresholds yielded noisy family-level estimates.Source: §A.8 · p. 19

B Returns by group

• The 200 accounts fell 1.24 points behind SPY by 9 October after leading by 0.42 points at the 5 October close.The accounts lost more than SPY during the late-period decline and recovered less on 9 October; group-mean divergence widened on 8 and 9 October.Source: B Returns by group · p. 19
• −1.23 points was the mean active return against SPY held at each account’s invested fraction; active returns were negative in all four families and 18 of 20 groups.The account mean return was +0.251%, compared with +1.478% for the exposure-matched benchmark.Source: B Returns by group · p. 19, Table 2 · p. 19, Figure 8 · p. 20

B.1 When the scorecard arrived

• Nonzero scorecards first reached groups on 4 or 5 October, so the competitive-versus-collaborative score display was active for only the final five of seven trading days.All scorecards delivered on 1 October showed zero returns, and all views admitting scores on 2 October withheld them.Source: §B.1 · p. 21

C.2 Outcomes and holdings

• DeepSeek and Opus showed regime-return spreads exceeding shuffled-account references; semiconductor holdings and exit timing distinguished some of the best and worst groups.Every Opus manager bought TSM with or without peers, while Opus collaborative groups later sold all their TSM at average prices of USD 465.6 and 473.6.Source: §C.2 · pp. 21–22
• Most overlap formed in the first two trading days: DeepSeek peer memory rose from 17% to 45%, and Grok’s board with pings from 24% to 54%.No isolated group exceeded 21% overlap at any close; some convergence also occurred without communication.Source: §C.2 · p. 22, Figure 9 · p. 23

C.3 Writing and wording

• On Opus’s board with pings, “verify independently” reached nine members; recurring headers and disclaimers also spread through groups.The first user of each cited recurring phrase had no peer use in view, while later adopters had a peer’s use in their delivered view.Source: §C.3 · p. 23, Figure 10 · p. 24

C.4 Trades

• Communicating groups traded in step more often than isolated groups in 15 of 16 cases at a 30-minute window, and their same-stock, same-direction first fills were closer in 14 of 16 groups.Within group-mornings, the estimated same-morning trade effect was 2.4 points, with a 95% interval from −4.1 to 9.3.Source: §C.4 · pp. 23–24

C.5 Evidence and effort

• Case reviews found qualifications dropped in relayed claims, including a yield range extended beyond its stated time boundary; one BAC error cannot carry full responsibility for sales.The BAC sellers cited other reasons, and BAC underperformed the benchmark on the relevant daily comparison.Source: §C.5 · pp. 24–25
• Managers’ research remained overlapping, and targeted requests were rare: the boards with pings made 18, with every request chain stopping at depth one.Requests naming a stock concerned one a peer had already researched that day in 71 to 95% of communicating-group cases.Source: §C.5 · p. 25

D Behavioural reviews

• Chronological reviews linked claims to delivered views, delegated research, and later actions; first-pass review findings retained quotation support after checks.The authors checked every text-cited case against source records, including orders, fills, posts, notes, and delivered views.Source: D Behavioural reviews · pp. 25–26

E Further cases

• Corrections did not always repair the analysis they addressed, and similar rule changes in isolated groups prevent attributing them to competition alone.A corrected oil-return comparison still contained an unchanged window; the recalculated full-window XLE–oil correlation was +0.498 rather than −0.103.Source: E Further cases · p. 26

F Operations and cost

• 91 wakes across 90 managers in 12 communicating groups failed before model execution on 1 October because peer context exceeded an obsolete readback limit.The issue was fixed after that day’s close; no isolated group was affected.Source: F Operations and cost · p. 26
• 105 of 106 self-wakes that expired without an attempt belonged to Sol, whose median delay to scheduled starts was 36.2 minutes.Median delays were 9.9 minutes for DeepSeek, 11.4 for Grok, and 14.3 for Opus; models could run at most nine cycles concurrently.Source: F Operations and cost · p. 26
• 1,633 research calls failed across 190 managers on the final day, with the cause unknown because provider error messages were not recorded.The failures affected all 20 groups; managers’ choices to proceed on weaker evidence are behavioral observations, while missing research is operational.Source: F Operations and cost · p. 26
• USD 7,259 of model inference was recorded over seven trading days: USD 5,130 for managers and USD 2,129 for delegated research.These are application estimates from token usage, not reconciled provider invoices; cost per group ranged from USD 13 to USD 852.Source: F Operations and cost · p. 26, Table 7 · p. 27

G Financial sensitivities and overlap

• Table 8 measures mean pairwise long-position overlap outside SPY as a share of account equity, averaged over seven closes and at the cutoff.Its 900 cutoff pair values reconcile to retained holdings, but accounts recur across pairs, so the values describe portfolio structure rather than independent regime-effect observations.Source: Table 8 · p. 27, G Financial sensitivities and overlap · p. 27
• Excluding two accounts with four pre-opening fills changes the overall mean return from 0.622% to 0.588%, a composition sensitivity rather than a same-account counterfactual.The secondary close-to-close interval yields a mean account return of +0.251%, against SPY’s +1.908%; the interval retains first-day holdings and histories and does not correct for the peer-context failure.Source: G Financial sensitivities and overlap · p. 27

H Behavioural evidence index

• Table 9 indexes sources behind the behavioral cases, retaining record references and chronology or calculation checks without adding independent observations.Coordinating review denotes direct checks against orders, fills, posts, memory publications, and delivered views without a group-review finding.Source: H Behavioural evidence index · p. 28, Table 9 · p. 29

I Channel settings

• Channel settings recorded on 8 October set upper bounds, not guarantees; delivered views capture what managers actually received, including the 1 October failure and late scores.Table 10’s limits concern manager-to-manager communication, not research.Source: I Channel settings · p. 28, Table 10 · p. 30

J Regime additions to the operator brief

• The six nonidentity operator-authored prompt sections were identical across all 200 agents, while the identity section had five regime-specific variants.Each communicating variant began with the complete isolated identity plus its regime addition; this comparison does not establish equality of complete runtime prompts across model families.Source: J Regime additions to the operator brief · p. 28

J.1 Competitive board; J.2 Collaborative board

• Competitive boards paired separate accounts with individual ranks and a shared message board, whereas collaborative boards used an equally weighted mean team score.Messaging became visible after a successful cycle; participants retained trading authority only over their own accounts.Source: §J.1 · p. 28, §J.2 · p. 28

J.3 Board with pings

• Boards with pings added structured requests that could trigger a peer’s additional cycle; ordinary @name text did not wake a peer.Pings and messaging were confined to the swarm and subject to scheduling and other platform limits.Source: §J.3 · p. 30

J.4 Peer memory

• Peer memory provided bounded, read-only snapshots of published peer notes, without a message board or agent pings.Snapshots were fixed for the current cycle and came from successfully completed cycles; each manager also had a bounded snapshot of its own configured memory when publication succeeded.Source: §J.4 · p. 30