More AI agents do not automatically yield a better system, study suggests

A recent Vals AI study tested agents that built 50 web apps from specifications and found that adding more agents does not reliably improve results. The main obstacle is coordination overhead, since more agents require managing communication, state synchronization, output merging, and conflict resolution. Like parallel computing, splitting work only saves time when the tasks can be divided effectively. In tightly coupled tasks, a single capable agent with greater reasoning effort or longer context can be cheaper and faster. The author argues that recognizing when not to use multiple agents is becoming a key design skill.