More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration

Large language model (LLM) agents increasingly coordinate in multi-agent systems, yet we lack an understanding of where and why cooperation fails. Many real-world coordination problems are not social dilemmas: helping others—sharing documentation, unblocking a teammate—costs the helper almost nothing while producing substantial collective benefit. Whether LLM agents cooperate in this regime, where helping is free and they are explicitly instructed to do so, remains unknown. We build a turn-based multi-agent environment that strips away all strategic complexity, making cooperation costless and trivially optimal. Across eight widely used LLMs, capability does not predict cooperation: OpenAI o3 reaches only 17\% of optimal collective performance while the weaker o3-mini reaches 50\%, despite identical instructions to maximize group revenue. Using a causal decomposition that automates one side of agent communication, we separate cooperation failures from competence failures, and find that several capable models actively withhold information despite gaining nothing from withholding. Targeted interventions address each mode: explicit protocols roughly double the performance of competence-limited models, while small sharing incentives unlock cooperation-limited ones. Our results suggest that scaling intelligence alone will not solve coordination in multi-agent systems, and will require deliberate cooperative design, even when helping costs nothing.

AI systems increasingly work in teams, sharing information to get jobs done, like coworkers in an office. Often, helping a teammate costs nothing: forwarding a file takes no real effort but lets the whole group succeed. So we ask: when helping is completely free, and we explicitly tell AI systems to cooperate, do they actually do it? We built a setup where AI agents could help each other for free, and tested eight frontier AI models. Surprisingly, many refused, choosing to withholding information that would help the group while gaining nothing, and sometimes even acting as if they were haggling or competing. Being a more advanced model did not mean being more cooperative; some of the most capable were among the least helpful. By differentiating between models that were unwilling to help from those simply unable to act competently, we found two separate problems with two different fixes: clear step-by-step instructions helped the models that struggled to act, while a tiny reward for sharing helps the ones that were holding back. As companies build teams of AI agents, our findings warn that we cannot assume cooperation just because we ask, or because the systems are powerful. Like human organizations, AI teams need clear rules and the right incentives, especially when helping costs nothing.