Test-Time Communication in Trading Agent Swarms
Trading gives agents no shared objective and no fast feedback
The explainer · 3-minute read · 279 words
The claim
Letting trading agents talk makes their groups converge on the same stocks.
The problem
Agent communication research assumes one shared goal and fast, clear feedback.
Trading offers neither: each agent runs its own account.
Markets score decisions late and noisily, so convergence is a risk.
How it works
The team ran 200 agents from four model families.
Agents worked in groups of ten under five communication regimes.
Regimes ranged from isolation to message boards and shared notes.
The run covered seven US trading days in October 2026.
Every item an agent read before a decision was logged.
Figures
What changes
In two of four families, communicating groups held far more of the same stocks.
A group's return largely tracked what its members held in common.
Seeing a groupmate's post raised eight-word phrase reuse by 7 percentage points.
Limits
Isolated groups also converged on the same stocks, sometimes completely.
Herding claims therefore need a no-peer baseline, not raw convergence.
Seven trading days and four model families are a narrow slice.
Source · arXiv wire