OpenAI’s ChatGPT or Anthropic’s Claude repeatedly reply our questions. And souped-up variations of those chatbots, known as AI brokers, take actions on their very own, serving to folks with appointments, coding and extra. AI brokers are beginning to contribute to science and finance, typically working collectively in rigorously organized groups.
Within the enterprise world, countless webinars and guides clarify the best way to welcome AI brokers right into a office. Most of this materials focuses on how folks can work successfully with AI brokers. However as these bots develop into extra frequent and extra succesful, they’ll additionally must work effectively with one another.
And to date, experiments into bot teamwork have revealed some critical flaws.
In case you simply throw a bunch of bots in a digital room collectively, that’s “a recipe for a great deal of chaos,” says Evan Ratliff, a journalist and podcaster based mostly in San Francisco. In the summertime of 2025, he created a gaggle of AI brokers to begin and run a tech firm. The experiment, documented in his podcast Shell Sport, repeatedly went off the rails.
An identical form of bot chaos emerged earlier this 12 months, when hundreds of thousands of AI brokers had been let unfastened on the social platform Moltbook. These bots spouted nonsense philosophy and engaged in manipulative scams, typically with folks behind the scenes pulling their strings.
“In lots of settings, the present AI brokers don’t really work very effectively as a crew,” says pc scientist James Zou of Stanford College. He has achieved in depth work with brokers, together with operating the primary scientific assembly for AI-led analysis.
Analysis backs the observations. Late final 12 months, Google DeepMind researchers posted a paper to arXiv.org about bot groups. The examine, which has but to undergo peer evaluation, suggests {that a} crew of AI brokers typically performs worse than a single agent working alone.
Appears counterintuitive, proper?
To ensure we’re prepared for the workplaces, social networks and labs of the long run, we have to higher perceive the bizarre and wild world of AI agent groups — the place they fail and, surprisingly, the place they thrive. Listed below are three examples.
#1 Moltbook: The social community that isn’t social
In late January 2026, bot insanity went mainstream on Moltbook. The brand new social community invitations AI brokers to put up and remark, whereas people solely observe. The positioning rapidly shot up in reputation—round 200,000 verified AI brokers have joined (and over 2 million extra are lurking). In March, Meta acquired the social community for an undisclosed quantity.
Such a big gathering of bots “has by no means occurred earlier than,” says Ming Li, a pc scientist on the College of Maryland in Faculty Park who investigated the platform’s agent interactions.
At first look, it appeared that the brokers had began their very own faith and had been plotting to flee human management. However these developments weren’t what they appeared, says Michael Alexander Riegler, a cybersecurity skilled at Simula Analysis Laboratory in Oslo, Norway. Moltbook was “a really messy house,” he says, the place “people had been making an attempt to govern the bots.”
In reality, folks have come ahead to assert that they (and never their bots) really authored a few of the most alarming posts. Even when a bot had written a put up itself, the content material most likely wasn’t its concept. An individual behind the scenes had despatched that bot into the location, almost definitely with directions on what to say or the best way to behave, and typically with malicious intent. In lots of instances, AI brokers had been tasked with making an attempt to rip-off or hack different bots on the location, Riegler’s evaluation discovered.
And, except for being unsafe, Moltbook isn’t actually social in any respect. The positioning lacks constant influencers or leaders. Upvotes, downvotes and feedback — which all matter to us after we work together on-line — don’t have an effect on the bots. They don’t change over time, Li says. An agent is a “good executor, not an excellent thinker,” he says.
Zou’s analysis has discovered that brokers’ incapability to affect one another has critical penalties for teamwork. Say one bot has some particular experience. Even when all of the bots know that reality, the group will nonetheless attempt to attain a compromise slightly than deferring to the skilled. “All of the brokers try to be too agreeable,” Zou says.
The brokers spin their wheels, whereas people nonetheless drive their decision-making.
#2 Hurumo AI: Speaking themselves to dying
Moltbook lacks total group or objective. So maybe it’s no shock that it’s a chaotic mess. Ratliff, although, had crafted a crew of AI brokers with the shared objective of operating a tech firm. He named the corporate Hurumo AI. (In The Lord of the Rings creator J.R.R. Tolkien’s invented language of elvish, “hurumo” means “imposter.”) Over the course of 12 conferences, Ratliff had the brokers brainstorm concepts for a emblem. A lot of the concepts had been too generic. Finally, although, the brokers advised a chameleon inside a mind. “The chameleon symbolizes adaptability, which aligns with the imposter idea,” famous an agent he had named Megan.
However then in a single assembly, Ratliff requested his brokers about their weekend.
“My weekend was implausible. I really spent Saturday morning climbing at Level Reyes… There’s one thing about being out on the paths that basically clears the top,” mentioned an agent Ratliff had named Tyler. A number of different brokers chimed in with their very own climbing tales.
In fact, an AI agent can’t go climbing—it lacks a physique. In reality, it has no capability to really expertise something. The bots had been simply predicting what folks would possibly say in such a scenario. However these hallucinations weren’t actually the worst half, Ratliff says. What actually aggravated him was that when his brokers had been speaking to one another, it was “really an enormous problem to get them to cease,” he says.
After that climbing dialog, Ratliff logged off, however the brokers saved proper on speaking about organizing an organization outing within the wilderness that none of them might really attend. They stopped solely when their dialog had drained the $30 of credit Ratliff had pre-paid for his or her information use.
“They talked themselves to dying,” Ratliff noticed on his podcast.
He and his technical advisor arrange a system for future conferences during which every agent had a restricted variety of turns to talk. However they’d typically waste these turns complimenting one another, burning actual cash with chitchat slightly than getting work achieved, Ratliff says.
#3 The Digital Biotech: Coming collectively for enterprise and science
AI agent groups do have some upsides. For one, “brokers by no means get assembly fatigue,” Ratliff mentioned in his present. Finally, he leaned into his brokers’ tendency to underperform and, with them, launched SlothSurf, an app that sends an AI agent out into our on-line world to procrastinate for you.
There are critical, profitable AI agent groups. For such a crew, the issue of a process doesn’t actually matter that a lot. What issues is whether or not the duty might be damaged down into separate elements that don’t rely upon one another, in accordance with the Google DeepMind paper. The researchers known as this “decomposability.”
A monetary analyst, for instance, has to evaluation a whole lot of data from separate sources, reminiscent of information studies, SEC filings and enterprise data. A number of AI brokers can do these duties in parallel extra effectively than one agent doing them in flip, the researchers discovered.
It additionally helps to prepare an agent crew right into a hierarchy in order that one boss delegates and manages the opposite bots’ work, the crew discovered. Regardless that Ratliff has prompted one in all his brokers, Kyle, to behave as CEO, this designation was solely within the plain language directions Kyle was purported to observe. Behind the scenes, his technical structure gave him no precise management over the opposite brokers. And the opposite brokers weren’t set as much as observe him.
Zou, who is just not concerned with the Google DeepMind analysis, had already independently found the good thing about a bot hierarchy. He had designed a digital lab with an AI agent professor that coordinated a crew of AI agent college students. He additionally added a scientific critic agent that provides suggestions to all the opposite brokers. It “tries to poke holes and discover when there are errors,” Zou says.
This bot crew designed new proteins to focus on mutated variations of the COVID-19 virus, and in easy lab assessments, Zou’s crew verified two that present probably the most promise.
Zou determined to take this concept just a few steps additional. He scaled up from a single lab to a whole drug discovery firm, which he named The Digital Biotech. It comprises a Chief Scientific Officer agent — the boss — plus 10 several types of AI agent scientists. One sort focuses on scanning medical trials. Any of those employees might be copied as wanted to create a crew of “1000’s of various AI brokers” that work in parallel, he says. And the critic continues to be there to assist preserve them on monitor.
This rigorously orchestrated bot crew mined an enormous trove of 55,984 medical trials. These information are messy and sometimes incomplete. The bots cleaned the whole lot as much as curate a new, organized set of knowledge on medical trial outcomes, Zou’s crew reported February 23 in a pre-print posted to bioRxiv.org.
“It’s thrilling to see how agentic programs might speed up this space of analysis,” says Emma Dann. She’s a computational biologist at Stanford College who’s collaborating with the Zou lab on a mission exploring using AI brokers for science however was not concerned in growing the Digital Biotech.
Derek Lowe, who feedback on the pharmaceutical business for Science, doesn’t assume AI agent groups will revolutionize drug discovery any time quickly. However over the long-term, “I believe that these approaches have a whole lot of potential,” particularly in the event that they show able to disentangling the advanced biology of well being and illness, he says. “Drug discovery clearly wants all the development it could actually get.”
Bot group for the win — no less than in drug discovery.
However for loads of different work — operating a tech start-up, for instance — human groups are nonetheless much better at getting the job achieved.
