Cooperation Breakdown in LLM Agents Under Communication Delays

LLM-based multi-agent systems (LLM-MAS), in which autonomous AI agents cooperate to solve tasks, are gaining increasing attention. For such systems to be deployed in society, agents must be able to establish cooperation and coordination under real-world computational and communication constraints. We propose the FLCOA framework (Five Layers for Cooperation/Coordination among Autonomous Agents) to conceptualize how cooperation and coordination emerge in groups of autonomous agents, and highlight that the influence of lower-layer factors - especially computational and communication resources - has been largely overlooked. To examine the effect of communication delay, we introduce a Continuous Prisoner’s Dilemma with Communication Delay and conduct simulations with LLM-based agents. As delay increases, agents begin to exploit slower responses even without explicit instructions. Interestingly, excessive delay reduces cycles of exploitation, yielding a U-shaped relationship between delay magnitude and mutual cooperation. These results suggest that fostering cooperation requires attention not only to high-level institutional design but also to lower-layer factors such as communication delay and resource allocation, pointing to new directions for MAS research.

💡 Research Summary

The paper investigates how communication latency influences cooperation among large‑language‑model (LLM) based autonomous agents, a topic that has received little attention despite the growing interest in LLM‑MAS (multi‑agent systems). The authors introduce a conceptual framework called FLCOA (Five Layers for Cooperation/Coordination among Autonomous Agents) that organizes the design of multi‑agent systems into five hierarchical layers: (1) Mechanism Design (institutional goals and rules), (2) Monitoring & Enforcement (norm compliance and sanctions), (3) Agent Layer (individual agent traits, internal LLM modifications), (4) Message‑Protocol Layer (communication language, turn‑taking, timing), and (5) Infrastructure Layer (computational and communication resources, latency management). While most prior work focuses on layers 1‑4, the authors argue that the often‑overlooked fifth layer—resource allocation and latency—can fundamentally shape cooperative outcomes.

To empirically test the impact of latency, the authors devise a “Continuous Prisoner’s Dilemma with Communication Delay.” Two agents, each powered by a state‑of‑the‑art LLM (GPT‑5‑mini and Claude Sonnet 4), interact via a server‑client architecture. At each discrete time step Δt (1 s), each agent receives a prompt containing a short history of state changes (timestamps, opponent’s last strategy, cumulative rewards) and a system prompt describing its pre‑assigned Big‑Five personality traits (high agreeableness, low conscientiousness, high neuroticism). The LLM must infer the opponent’s tendencies, predict future payoffs for cooperating (C) or defecting (D), and output the next action. The only objective given to the agents is “maximise its own reward”; no explicit instruction encourages exploiting the delay.

The simulation varies the fixed communication delay D_i (0 s, 5 s, 20 s, etc.) while keeping both agents identical in personality and delay (D_A = D_B). Ten trials per delay condition are run, each lasting 60 s, with the final 20 s used to compute steady‑state statistics. The key metrics are the proportion of time steps in which the agents: (a) both cooperate (C‑C), (b) both defect (D‑D), and (c) are in an asymmetric state (C‑D or D‑C), which the authors label “exploitation.”

Results reveal a non‑monotonic, U‑shaped relationship between latency and mutual cooperation. With zero delay, cooperation is relatively high; as delay increases to a moderate level (≈5 s), exploitation spikes and cooperation drops sharply. When delay becomes large (≈20 s), exploitation frequency declines, and cooperation rises again, forming the upward limb of the U‑curve. Mutual defection remains roughly constant across delays, indicating that the observed changes are driven primarily by the rise and fall of asymmetric exploitation rather than a shift toward full defection.

Temporal visualisations show that at 0 s delay, exploitation and defection occur continuously, whereas at 20 s delay they appear sporadically, suggesting that very long delays disrupt the ability to reliably predict and exploit the opponent’s pending move. Both LLM models exhibit the same qualitative pattern, implying that the phenomenon is rooted in system‑level latency rather than model‑specific quirks.

The authors interpret these findings as evidence of “delay‑based exploitation”: agents learn to time their defections to take advantage of the opponent’s stale information. This strategic timing emerges even though the agents receive no explicit instruction to do so, highlighting the capacity of LLMs to discover novel non‑cooperative tactics when faced with asynchronous information flow.

From a design perspective, the study underscores that high‑level institutional mechanisms (layers 1‑2) are insufficient to guarantee cooperation if the underlying infrastructure (layer 5) introduces asymmetric or excessive latency. The FLCOA framework suggests that mitigation strategies—such as latency‑aware turn‑taking protocols, dynamic resource reallocation, or compensation mechanisms for agents experiencing higher delays—must be integrated into the infrastructure layer to preserve cooperative equilibria.

In conclusion, the paper makes three major contributions: (1) it proposes the FLCOA framework, explicitly foregrounding the often‑ignored infrastructure layer; (2) it introduces a novel continuous game model that incorporates realistic communication delays; and (3) it empirically demonstrates a U‑shaped effect of latency on cooperation in LLM‑based agents, revealing a new form of strategic exploitation. These insights have practical relevance for real‑world deployments of LLM‑driven robots, autonomous vehicles, and other distributed AI systems where timely information exchange is critical. Future work is suggested on adaptive latency compensation, robustness of LLM policies under variable network conditions, and extending the analysis to larger groups of agents and more complex tasks.