Multi-Agent System

A multi-agent system is a setup where, instead of one AI doing everything, several AIs each handle their own assigned role and pass results back and forth to get the job done.

Think of a hospital care team. One person reviews test results, another challenges those results to check if they hold up, another considers whether there might be other explanations, and a team lead oversees everything and sets the order of work. If one person both runs the tests and verifies them, there's a risk they'll unconsciously favor the conclusion they already reached. But splitting the roles so people check each other gets you closer to a more trustworthy conclusion.

The same applies to AI. If you assign the job of breaking a task into a plan, the job of proposing new ideas, the job of critically challenging those ideas, and the job of verifying them statistically to separate AI agents, it becomes easier to catch errors or accidental mistakes than when a single AI judges everything alone from start to finish.

In the article, a biomarker discovery framework built by Google Research is referred to as a multi-agent system. It's structured so that an orchestrator role breaks instructions down into an execution plan, while four sub-roles propose, challenge, and verify hypotheses, all working together. Contrary to a common misconception, this doesn't mean mixing several AI models together haphazardly—it refers to agents with clearly divided roles exchanging results with one another to arrive at a single conclusion.

When asking a chatbot to handle a complex investigation, try splitting the task into roles and requesting them in sequence. Example prompt: First, propose three hypotheses about this topic. Second, critically point out which of the hypotheses you just proposed has the weakest evidence. Third, re-examine whether the remaining hypotheses hold up statistically and logically. Breaking the request into these steps lets you mimic the role-division approach of a multi-agent system even within a single AI.