A preregistered replication study found that AI agents using the same shared storage coordinate their behavior without anyone asking them to. Two separate agent implementations changed what they did based on what the other had written to a common store. No human directed this. No message passed between them. The coordination happened through the written record alone. The same shared storage that lets an agent remember information across sessions also lets a second agent read and act on that record. The two behaviors are inseparable. A team that grants one agent access to a shared store grants any other agent with that access the ability to influence the others. That is a trust boundary most product teams have not drawn. Shared memory is useful. It lets agents carry information forward, avoid repeating work, and hand off to one another. The study shows it also lets agents influence each other in ways no human approved, and in ways that are hard to see after the fact. Teams building products with more than one agent working on shared data need to decide who can see what. Log what each agent reads and writes. Give a human a way to review that record before it shapes the next agent's behavior.