Complex information-seeking tasks often require multiple AI agents working together. A central agent delegates subtasks to specialized agents, but coordinating them is difficult: agents may receive poorly defined tasks, duplicate work, or share too little context, causing errors that are hard to diagnose and repair manually.
This paper tests an automated method called MultiLoop that uses a coding program to analyze execution traces, identify failures, and rewrite the agents' prompts, tools, and coordination logic. The method was applied to a multi-agent system on the DeepSearchQA benchmark using only ten training examples.
The evolved system improved accuracy by 4.6% over the original multi-agent setup and by 3.2% over the best single-agent system. However, the underlying coordination problems, such as poor task delegation and insufficient information sharing, remained largely unchanged. Future work must focus on optimizing how agents interact, not just their individual behavior.