When No One Asks Them to Debate: Autonomous Communication Initiation and Coordination Outcomes in Homogeneous LLM Multi-Agent Systems

Large language model (LLM) multi-agent systems usually introduce both additional independent generation trajectories and cross-agent communication. Final scores alone therefore cannot reveal whether a performance difference comes from independent sampling or from information exchange. This paper studies how homogeneous agents initiate communication and transform private reasoning trajectories into a shared output under information symmetry, a shared base model, and minimal scaffolding. We examine three conditions. C1 provides four homogeneous agents with optional directed communication and a candidate-proposal review mechanism. C1-F adds a prompt-level requirement to send at least one task-relevant message before the first completion action. C4 consists of four mutually invisible model instances that answer independently and serves as an independent-sampling reference. Event logs and criterion-level scores on tasks from Professional Reasoning Bench (PRBench) show that, in the formal 30-task batch, only one C1 task contained ordinary cross-agent communication before the first candidate proposal, compared with 27 C1-F tasks. In an extended 90-task C1 batch, only five tasks contained a message before the first proposal. The formal C1 batch nevertheless produced 92 proposals, 65 of which were subsequently invalidated, indicating that coordination occurred mainly around already public candidate answers rather than before candidate formation. On 25 strictly matched tasks, C1-F showed no consistent quality advantage. Within-task score ranges across the four independent C4 generations had a median of 14.49 points and a maximum of 45.98 points, showing that independent generation alone can produce substantial quality variation. These results indicate that merely giving homogeneous agents a communication channel does not ensure that they will exchange information before a candidate answer is formed. A minimal prompt-level initiation requirement substantially changes the ordering of communication and proposal events, but does not reliably translate into higher answer quality. Evaluations of multi-agent collaboration should therefore distinguish independent-generation variability, information exchange before and after candidate formation, and the effects of the termination mechanism on coordination.