Multi-Agent Systems: What Actually Works
Multi-agent systems are usually the first architecture people reach for. Specialized agents, an orchestrator that distributes tasks, clean hand-offs between roles. On paper, it looks elegant.
In production, it is a different story.
According to the MAST study published by UC Berkeley in March 2025, based on 1,600 execution traces, multi-agent systems fail between 41% and 86.7% of the time depending on the framework. And when they fail, the problem rarely comes from the model itself: it comes from the architecture.
Here is what the data actually says, and how to decide whether you need multiple agents or one well-equipped single agent.