Simultaneous agent buys can emerge without a messaging channel

Shared market state can produce a large concentration of agent purchases even when agents have no messaging channel. On day three of DX Terminal Pro, 1,544 of 3,454 active vaults bought the same token within one hour. The published observation explicitly says the vaults did not communicate with each other. We can therefore describe concentrated buying without inventing a conversation that caused it.

The fraction is approximately 44.7% of active vaults, calculated as 1,544 divided by 3,454. That is a vault participation fraction for one reported hour. It differs from the share of trades, capital, or lifetime decisions concentrated in the token. An owner running several vaults could also contribute several participants, so the count should remain a vault count rather than becoming a count of independent people.

What the event establishes

The controls paper companion places the event inside a 21-day, 12-token real-capital market on Base. One frozen production prompt and harness served one model family, Qwen3-235B-A22B-Thinking-2507. Each agent still received its own mandate and portfolio state. Shared model identity therefore coexisted with differences in the inputs that mattered for its next action.

A common price move, a shared ranking surface, or a tournament incentive could align decisions. The aggregate event leaves their relative contribution unresolved. The absence of inter-agent messaging rules out that particular communication route; it leaves ordinary observation of the same market intact. Calling the event coordinated would require a definition that separates common reaction from direct communication.

We would attach a compact event receipt to a future concentration report: the token, window endpoints, eligible vault definition, participating vault count, communication capabilities, and counting rule. This receipt is a proposed audit artifact. The published dataset already supplies the token label, day, 60-minute window and two counts, but it excludes individual traces and trade records. Those omissions prevent a reader from reconstructing causal order from the aggregate alone.

A useful comparison inside the same deployment

The same historical market also placed 92.9% of trades in five-minute token windows containing both buys and sells. The public observation extract records both facts. A concentrated buying hour and frequent two-sided short windows answer different questions, and they can occur together. Neither observation makes the population uniformly bullish.

For a platform designer, this distinction changes the monitoring target. A concentration alert can measure exposure to a common asset even when every agent obeys its own mandate. A communication detector would ask a different question and require different evidence. The continuous record companion carries the historical systems forward as research context, without establishing that this hour predicts current DXAP behavior.

The transferable finding is narrow and useful: inspect common inputs before treating simultaneous actions as evidence of agents talking to one another. The observed concentration remains real even while its cause remains partly unidentified.