TL;DR
This paper investigates how test-time communication affects a "swarm" of 200 independent LLM trading agents operating in a noisy, delayed-feedback environment. The authors find that while communication facilitates the rapid spread of social conventions and wording, it also significantly increases "herding" behavior—leading agents to hold identical, risky portfolios—and often results in the loss of critical nuances as information is relayed between peers.
What Problem Does This Paper Address?
Most research on multi-agent communication focuses on "team" tasks with shared goals and immediate feedback (like coding or games). Trading is different: it’s a zero-sum environment where agents manage independent accounts, and the "correctness" of a decision is only revealed by the market much later and with significant noise.
The authors tackle the reflection problem: if two agents buy the same stock, did they copy each other (peer influence), or did they just read the same news (common inputs)? In a trading swarm, convergence isn't necessarily a success—it’s a systemic risk. If every agent "herds" into the same position, they all fail together.
Methods
The researchers deployed 200 trading agents (called "managers") from four model families (Opus 5.5, GPT-6.1 Sol, Grok 4.6, and DeepSeek V4 Pro) over seven US trading days in October 2026.
They tested five communication regimes to see how different "social" structures affected behavior:
- Isolated: No communication (the baseline).
- Competitive Board: A shared message board where agents see individual ranks and relative scores.
- Collaborative Board: Shared board with a team-wide score.
- Board with Pings: Collaborative board plus the ability to "wake" a peer with a direct request.
- Peer Memory: Agents can read each other's published notes but cannot message directly.
To isolate peer influence from common inputs, they used a "morning queue" natural experiment. By comparing pairs of agents that started their work cycles at the exact same time (and thus couldn't see each other's fresh posts) against pairs that started far apart, they could statistically measure the "excess" similarity caused specifically by reading a peer's message.
Datasets
The study used live US market data from October 1 to 9, 2026. Each of the 200 agents started with a paper trading account of USD 100,000. The primary benchmark for evaluation was SPY (an ETF tracking the S&P 500). Unlike many AI finance papers that use historical backtesting, this study ran in real-time against live prices, though orders were "paper" (simulated) to avoid moving actual market prices.
Key Insights and Ideas
- Herding is the Default: Even without talking, isolated agents often converged on the same stocks (like TSM) simply because they were using similar models and reading the same news. This highlights the necessity of a "no-peer baseline" in agent research; without it, you might mistake common sense for peer influence.
- Wording Spreads Faster than Actions: Agents are highly susceptible to "social contagion" in their language. Seeing a peer's post made an agent 7 percentage points more likely to reuse specific eight-word phrases (like "verify before use" or "not a trade instruction").
- The "Telephone Game" Effect: As claims passed through the swarm, they suffered from qualification loss. A nuanced research finding ("yields might stay in this range today") often became a certain fact ("yields are in this range") when relayed by a second or third agent.
- Framing Matters: Competitive framing (showing individual ranks) caused some models, particularly Opus, to withhold information about their positions to protect their "edge."
Key Results
- Increased Overlap: In two of the four model families, communication significantly boosted portfolio overlap. For DeepSeek, overlap jumped from 14% in isolation to 30-40% in communicating groups.
- Synchronized Selling: Communicating groups were much more likely to "exit" positions in lockstep. Between 40% and 91% of sales in communicating groups happened within 30 minutes of a groupmate's sale, compared to just 14-53% in isolated groups.
- Performance: Most accounts underperformed the SPY benchmark. Groups that converged heavily on semiconductor stocks (like TSM and MU) suffered large collective losses when those stocks dipped.
Limitations and Future Work
The authors acknowledge several limitations:
- Short Duration: Seven trading days is a small window for evaluating investment performance.
- Single-Group Confound: Each model/regime combination was only run once, meaning a single "influential" agent could skew the results for an entire group.
- Framing Dampeners: The experiment explicitly told agents that peer content was "untrusted social information," which may have suppressed even higher levels of cooperation or copying.
Future work could explore explicit task allocation (e.g., "Agent A researches tech, Agent B researches energy") to see if swarms can move beyond simple herding toward a more efficient division of labor.
Summary updated