Autonomous AI agents for cloud infrastructure management — including incident remediation, cost optimization, and capacity planning — have gained significant attention in recent years as operational complexity surpasses human capacity for real-time decision-making. However, deploying multiple autonomous agents on shared infrastructure introduces coordination challenges such as conflicting actions, resource contention, and emergent behaviors that are difficult to predict analytically. Conventional approaches predominantly depend on centralized orchestration or static rule-based controls, which are often inadequate for handling dynamic multi-agent interactions at scale. To overcome these limitations, this paper proposes MACOS (Multi-Agent Cloud Operations Simulator), a discrete-event simulation framework driven by Monte Carlo methods that models the behavior of interacting autonomous agents across large-scale cloud environments. The framework introduces a novel Coordination Entropy Metric (CEM) that quantifies the degree of coordination breakdown between agents, enabling operators to identify dangerous interaction patterns before production deployment. Four coordination protocols — centralized orchestration, token-based mutual exclusion, consensus-based negotiation, and stigmergic coordination inspired by biological swarm systems — are evaluated under varying failure scenarios and agent densities. Experimental results demonstrate that uncoordinated multi-agent operations exhibit a 34% conflict rate at scale, while stigmergic coordination reduces conflicts to under 3% with minimal communication overhead. The study further reveals a phase transition phenomenon where system stability degrades sharply beyond a critical agent density threshold. This work highlights the importance of embedding simulation-based coordination testing and entropy-aware monitoring into autonomous cloud operations to enable safe, self-adaptive infrastructure management for large-scale cloud-native ecosystems.