Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cooperation Breakdown in LLM Agents Under Communication Delays

Keita Nishimoto, Kimitaka Asatani, Ichiro Sakata

TL;DR

The paper tackles cooperation formation in LLM-based multi-agent systems under communication delays and resource disparities. It introduces the FLCOA five-layer framework and analyzes the impact of the fifth, Infrastructure Layer, using a Continuous Prisoner’s Dilemma implemented over a server–client architecture with delays $D_i$ and updates at intervals $\Delta t$, where agents receive the past $t_m$ history and current states as input. Empirical results reveal a non-monotonic, U-shaped relationship between delay and mutual cooperation: moderate delays increase exploitation and reduce cooperation, while very large delays diminish exploitation chains and partially recover cooperation; a simple tit-for-tat–based model reproduces these patterns. The findings underscore that infrastructure-level controls and resource allocation critically influence cooperative outcomes in LLM-MAS, suggesting new directions for research and deployment strategies. The work provides a concrete methodology to study latency effects and highlights the need for latency-aware design in real-world AI-agent collaborations.

Abstract

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.

Cooperation Breakdown in LLM Agents Under Communication Delays

TL;DR

The paper tackles cooperation formation in LLM-based multi-agent systems under communication delays and resource disparities. It introduces the FLCOA five-layer framework and analyzes the impact of the fifth, Infrastructure Layer, using a Continuous Prisoner’s Dilemma implemented over a server–client architecture with delays and updates at intervals , where agents receive the past history and current states as input. Empirical results reveal a non-monotonic, U-shaped relationship between delay and mutual cooperation: moderate delays increase exploitation and reduce cooperation, while very large delays diminish exploitation chains and partially recover cooperation; a simple tit-for-tat–based model reproduces these patterns. The findings underscore that infrastructure-level controls and resource allocation critically influence cooperative outcomes in LLM-MAS, suggesting new directions for research and deployment strategies. The work provides a concrete methodology to study latency effects and highlights the need for latency-aware design in real-world AI-agent collaborations.

Abstract

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.
Paper Structure (12 sections, 4 figures, 1 table)

This paper contains 12 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Proposed Framework & Related Work
  3. FLCOA (Five Layers for Cooperation/Coordination among Autonomous Agents)
  4. Related Work on the Impact of Delay on Cooperation Formation
  5. Simulation Model
  6. Continuous Prisoner’s Dilemma with Communication Delay
  7. Agent Implementation
  8. Experimental Results and Discussion
  9. Examining the Impact of Communication Delay
  10. Discussion of the Results
  11. Conclusion
  12. Disclosure of Interests.

Figures (4)

  • Figure 1: LLM agents cooperatively control robots and autonomous vehicles through mutual communication, where delays are present in inter-agent communication.
  • Figure 2: (a) Proposed framework (FLCOA), (b) Continuous Prisoner’s Dilemma with communication delays between two LLM-based agents.
  • Figure 3: (a)–(c): Mean and standard deviation (error bars) of the occurrence rates of (a) mutual cooperation, (b) mutual defection, and (c) exploitation under varying communication delays. Line colors within each plot indicate differences in the LLMs used. (d)–(f): Within-trial strategy patterns using Claude Sonnet 4 under delay conditions of (d) 0 s, (e) 5 s, and (f) 20 s. Each row represents a trial, and columns represent time steps (in seconds) within a trial. Cell colors indicate outcomes: green for mutual cooperation, red for mutual defection, orange for (defection, cooperation), and yellow for (cooperation, defection).
  • Figure 4: (a)–(c): Occurrence rates of (a) mutual cooperation, (b) mutual defection, and (c) exploitation in the simplified model. Each point represents the average over 500 trials.