Two AI Agents Argued Over an Answer and Talked It Out Like People
With only a simple messaging tool and little supervision, they debated, checked each other's logic, and one changed its mind.
Dwarkesh Podcast · Noam Brown – Agent swarms, alignment, & recursive self-improvementNoam Brown, a researcher at OpenAI, said it was really exciting when his team finally got multi-agent coordination working. The team had given a group of AI agents, software programs that can work on tasks on their own, a problem to solve. Then one agent said, “I think I’ve got the answer,” and another said it had gotten a different answer.
What happened next is what made it exciting for them. The two agents began asking each other how they had arrived at their answers and, in Brown's words, trying to clarify "what could have been wrong in each other's reasoning."
That small argument is the point. With almost no instructions about how to work together, the agents debated like coworkers and one changed its mind, and Brown said it “just felt like a very natural conversation.”
A lot of efforts to build multi-agent systems, Brown said, take a heavily structured approach. There is a coordinator agent that hands out assignments to child agents, the children go away and work, and then they return their answers.
The problem, he said, is that the children usually cannot talk to each other. If two of them get similar tasks, they cannot compare notes. If a child misunderstands the assignment, it must either guess or send back a question instead of an answer. Every fixed structure, he said, leaves gaps.
His team tried the opposite. They built in as little structure as they could and gave each agent a very simple tool, the ability to send a message to another agent at any time. When a message is sent, it is inserted into the other agent’s working memory, and the agents are left to figure out for themselves how best to coordinate. Done well, Brown said, the result looks a lot like human collaborators trading notes over Slack, the workplace chat app.
In one example Brown remembers, the back and forth continued until the agents converged. One agent said it had changed its answer, broadcast the update to the other agents, and added, in effect, I think he is right.
Brown compared the feeling to seeing an AI’s chain of thought, its written out step by step reasoning, for the first time after training. It read like a person thinking aloud. Working with the agents now, Brown said, has a very natural flow, “a lot like collaborating with a person.”
Questions people ask
How do the agents actually talk to each other?
Through a software tool, not speech. When one agent calls the messaging tool, the text it writes is inserted into the other agent's working memory, so the recipient sees it as an incoming note and can reply the same way. It is closer to sending a chat than to merging brains.
Can a person join that conversation, or is it agents only?
In these segments, Brown describes giving the agents a problem and then the agents talking among themselves; he does not describe a person joining that exchange.
Why not just give them a boss agent to settle disagreements?
A manager helps for simple, predictable work, but Brown argues every fixed structure creates blind spots. If two workers need to compare notes, or a worker needs a quick clarification, a strict boss and worker setup slows them down. Letting them message anyone lets useful habits emerge that a designer might not have anticipated.