Cheap Talk Enables Collective Repair in LLM Societies

Large-language-model (LLM) agents are increasingly deployed in autonomous settings, creating multiagent environments where numerous LLM agents communicate, negotiate and jointly allocate resources with minimal human supervision. Yet it remains unclear to what extent these agents' verbalized commitments govern their actions. To investigate how cheap talk affects coordination among agents, we set up two public-goods environments: a linear game in which every marginal contribution is strictly dominated in a one-shot interaction, and a threshold game in which a provision point must be met and donations are not refunded when provision fails. We introduce four perturbations: availability of cheap talk via reciprocal chat, number of players, probability of continuation, and exposure to an instructed defector mid-game. Our main observation is emergent collective repair following an instructed defection. After one instructed defector makes an established cooperation convention infeasible, two model variants consistently derive and sustain an alternative convention that compensates for that agent's defection (10 of 10 runs), while others fail to consistently sustain cooperation. No silent control sustains repair for any model variant (0 of 40 runs across eight variants). Cheap talk also sustains coordination beyond dyads. For pledge-abiding models, cheap talk supports one-shot commitments without future enforcement, while other models produce fluent agreements that fail to translate into action. Such behavior suggests that, when equipped with a channel for communication, LLM collectives can not only establish cooperative conventions but actively update and repair them through decentralized renegotiation rather than externally imposed rules. Our results position cheap talk as a powerful but model-dependent tool for forming, maintaining, and revising collective conventions.