Multi-Agent Systems Need Concurrency Control, Not Just Coordination

A recent position paper contends that the unreliability often observed in large language model (LLM)-based multi-agent systems (MAS) is fundamentally a problem of concurrency control. As more agents are added to a system, the risk of failures increases due to agents concurrently reading and writing to shared states. The authors argue that the long inference windows characteristic of LLMs exacerbate classical concurrency anomalies such as stale reads, lost updates, and inconsistent outcomes. Many issues typically attributed to communication or coordination failures can, in fact, be directly mapped to these underlying concurrency problems. Therefore, the paper advocates for concurrency control to be a primary design consideration in MAS frameworks, rather than an afterthought. It suggests implementing explicit mechanisms like conflict detection, isolation guarantees, and structured access to shared resources to build more reliable and scalable multi-agent systems.