How heterogeneous agents coordinate through a shared market

Agents can produce concentrated activity through shared information even when they exchange no messages. We propose testing how heterogeneous mandates change that concentration, then separating common-input reaction from communication and price feedback. This is a proposed experiment; the historical aggregate provides motivation rather than a result for the new design.

The public market observations record a one-hour event in which 1,544 of 3,454 active vaults bought the same token and describe the vaults as noncommunicating. The share is about 44.70%. The controls paper companion bounds those observations to a frozen 21-day deployment in a twelve-token market. That population cannot identify the effect of adding communication to a new fleet.

Change one interaction channel at a time

Construct three offline environments with identical initial portfolios, mandates and saved market observations. In the first, agents receive a common state but their simulated actions cannot alter later observations. In the second, agents can exchange structured research messages while the price path remains fixed. In the third, actions enter a declared market simulator and change later state, with communication again disabled.

These arms distinguish three mechanisms: common stimulus, direct information exchange and market-mediated interaction. A fourth combined arm can follow once each component is understood. Randomize agent-to-mandate assignments, keeping the distribution of constraints constant across arms. A common model version should remain fixed so differences follow from interaction channels.

An illustrative concentration statistic counts how many of 100 agents add exposure to the same asset within a window. If forty do so, participation concentration is 40%. It says little about size: forty $10 purchases total $400, while four $1,000 purchases total $4,000. Report participant count and signed notional together. Also inspect whether the agents hold correlated exposures before the event; a common exit can arise from shared holdings rather than shared new information.

What would change the interpretation

Score mandate compliance, asset concentration, turnover and adverse simulated price feedback. Communication could reduce duplicated research while increasing correlated action. Heterogeneous mandates could distribute exposure while introducing conflicting demands on shared liquidity. Both effects require measured evidence within the chosen simulator.

The continuous record companion shows heterogeneous outcomes and two-sided activity in historical systems, but it preserves a directional-edge null. We would keep that limitation attached when motivating fleet experiments. A more coordinated population can still lose after costs, and greater disagreement can still produce harmful churn.

The proposed deliverable is an interaction matrix with explicit channels and simulator assumptions. It would say which source of coupling generated each observed concentration change. That is a more useful basis for deciding whether agents should share research than labeling simultaneous activity as coordination from a timestamp plot alone.