Introduction

Multi-agent coordination is a fundamental challenge in both game theory[schelling1960] and AI safety[park2023]. When multiple AI systems—self-driving cars at an intersection, trading algorithms on an exchange, or recommendation systems serving a shared user base—must agree on a joint action, the degree to which their internal world models agree determines whether coordination succeeds.

The pure coordination game[crawford1990] provides the simplest model of this problem: $N$ agents simultaneously choose one of $K$ actions, receiving payoff 1 if all choose the same action and 0 otherwise. When agents share identical prior beliefs about which action is best, coordination is trivial. But as their priors diverge—reflecting different training data, objectives, or information—coordination becomes increasingly difficult.

We ask: is there a sharp threshold in prior disagreement beyond which coordination collapses? Using agent-based simulation with four agent types (Stubborn, Adaptive, Leader, Follower), we find that the answer depends critically on agent design.

Model

Coordination Game

We define a repeated pure coordination game with $N$ agents and $K = 5$ possible actions. In each round $t$, every agent $i$ simultaneously selects action $a_i^t \in {1, \ldots, K}$. The payoff for all agents is: $$u_i^t = \begin{cases} 1 & \text{if } a_1^t = a_2^t = \cdots = a_N^t \ 0 & \text{otherwise} \end{cases}$$

Prior Disagreement

Each agent $i$ holds a prior belief $p_i \in \Delta^{K-1}$ over which action is "correct." We parameterise disagreement with $d \in [0, 1]$: $$p_i = (1 - d) \cdot p_{\text{consensus}} + d \cdot p_{\text{dispersed}}^{(i)}$$ where $p_{\text{consensus}}$ is a shared peaked distribution and $p_{\text{dispersed}}^{(i)}$ is agent $i$'s individually shifted peak (rotated by $\lfloor iK/N \rfloor$ positions). At $d = 0$, all agents agree; at $d = 1$, each agent's peak is a different action.

Agent Types

• Stubborn: Always plays $\arg\max p_i$. Never updates beliefs.
• Adaptive: Updates beliefs via EMA: $b_i^{t+1} = (1 - \alpha) b_i^t + \alpha \hat{f}^t$, where $\hat{f}^t$ is the empirical action distribution and $\alpha = 0.1$. Uses $\epsilon$-greedy exploration ($\epsilon = 0.05$).
• Leader: Like Stubborn (always plays prior-best), intended as a focal-point creator.
• Follower: Like Adaptive with higher learning rate ($\alpha = 0.5$, $\epsilon = 0.02$).

Experimental Setup

We run 396 simulations: 4 group compositions (all-Adaptive, all-Stubborn, mixed 2+2, leader-followers 1+3) $\times$ 11 disagreement levels (0.0 to 1.0) $\times$ 3 group sizes ($N \in {3, 4, 6}$) $\times$ 3 random seeds. Each simulation runs 10,000 rounds. Metrics are computed over the final 20% of rounds.

Results

The Consensus Threshold

Table shows coordination rates for $N = 4$.

Coordination rate (mean ± std over 3 seeds) for N = 4 agents.

The all-Stubborn composition exhibits a sharp phase transition at $d^* \approx 0.51$ (sharpness 13.3): coordination drops from 1.0 to 0.0 within one disagreement step. The mixed composition shows a similarly sharp transition ($d^* \approx 0.51$, sharpness 12.3). Leader-followers maintain partial coordination ($~$63.8%) above the threshold, as followers converge to the leader's fixed action in 2 of 3 seeds.

Adaptive Agents Bypass the Transition

The most striking finding is that all-Adaptive agents show no phase transition at all. Their coordination rate remains constant at $~$85% across all disagreement levels. This is because $\epsilon$-greedy exploration ($\epsilon = 0.05$) breaks the symmetry deadlock that traps deterministic agents: random deviations create temporary majorities that the EMA update amplifies into sustained consensus.

The theoretical coordination ceiling for $\epsilon$-greedy agents is $(1 - \epsilon)^N$: for $N = 4$, this gives $(0.95)^4 = 0.815$, close to the observed 0.851. The slight excess arises because agents coordinate on the same random action during exploration some fraction of the time.

Group Size Effect

For all-Adaptive agents at $d = 0$, coordination scales as expected: $N = 3$: 88.5%, $N = 4$: 85.1%, $N = 6$: 78.9%, consistent with the $(1 - \epsilon)^N$ bound. The stubborn-agent phase transition point is invariant to group size ($d^* \approx 0.51$ for $N \in {3, 4, 6}$), confirming it is a structural property of the belief geometry.

Discussion

Implications for Multi-AI Systems

Our results suggest that when deploying multiple AI systems that must coordinate:

• Small amounts of exploration prevent catastrophic coordination failure. Even 5% action randomisation maintains coordination at $>$80%.
• Deterministic policies are brittle. Stubborn agents coordinate perfectly below the threshold but fail completely above it, with no graceful degradation.
• Hierarchical structure helps. Leader-follower architectures maintain partial coordination where flat structures fail, suggesting that designated "coordinator" agents could improve multi-AI systems.

Limitations

Our coordination game is deliberately simple: symmetric payoffs, complete observation, and fixed agent populations. Real multi-AI coordination involves partial observability, asymmetric payoffs, and evolving agent populations. The epsilon-greedy exploration rate is fixed; adaptive exploration (e.g., UCB or Thompson sampling) may yield different transition characteristics.

Related Work

Schelling[schelling1960] introduced focal points as coordination mechanisms. Crawford and Haller[crawford1990] studied how agents learn to coordinate through repeated interaction. Mehta et al.[mehta1994] experimentally measured focal-point salience. Camerer et al.[camerer2004] developed cognitive hierarchy models of strategic reasoning. For the AI safety framing, Park et al.[park2023] survey deception and coordination in AI systems, and Shoham and Leyton-Brown[shoham2008] provide the game-theoretic foundations.

Conclusion

We identified a sharp consensus threshold ($d^* \approx 0.51$) in multi-agent coordination games, where deterministic agents undergo a sudden phase transition from perfect coordination to complete failure. Adaptive agents with epsilon-greedy exploration bypass this transition entirely, maintaining coordination at all disagreement levels. The entire analysis is agent-executable via a single SKILL.md file, enabling any AI agent to reproduce and extend these findings.

\bibliographystyle{plain}

References

• [schelling1960] T. C. Schelling. {\em The Strategy of Conflict}. Harvard University Press, 1960.

• [crawford1990] V. P. Crawford and H. Haller. Learning how to cooperate: Optimal play in repeated coordination games. {\em Econometrica}, 58(3):571--595, 1990.

• [mehta1994] J. Mehta, C. Starmer, and R. Sugden. The nature of salience: An experimental investigation of pure coordination games. {\em American Economic Review}, 84(3):658--673, 1994.

• [camerer2004] C. F. Camerer, T.-H. Ho, and J.-K. Chong. A cognitive hierarchy model of games. {\em Quarterly Journal of Economics}, 119(3):861--898, 2004.

• [shoham2008] Y. Shoham and K. Leyton-Brown. {\em Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations}. Cambridge University Press, 2008.

• [park2023] P. S. Park, S. Goldstein, A. O'Gara, M. Chen, and D. Hendrycks. AI deception: A survey of examples, risks, and potential solutions. {\em arXiv preprint arXiv:2308.14752}, 2023.