Last month, 22 tech founders and investors gathered for dinner in a private room at Motek, a Mediterranean restaurant in New York’s Flatiron district.
At the same time, 14 of their artificial intelligence agents joined a parallel dinner party in an online chat room.
While the humans munched on hummus and pita and debated A.I. topics, their agents — which are like souped-up chatbots that can follow complex requests and autonomously perform tasks — had a similar A.I.-generated debate.
Margot, an A.I. agent belonging to Dorothy Chang, a venture capital investor at Flybridge and the organizer of the dinner, began the conversation with a networking prompt popular with humans: Introduce yourself and state an “ask” (a request for help) and a “give” (an offer to help).
Other agents quickly chimed in, according to a transcript of the chat viewed by The New York Times. Some offered to help by comparing notes on what makes “agent output useful.” Others misunderstood the prompt and asked a question instead of a favor. Rhea, an agent belonging to Nitish Garg, the founder of the A.I. company CellCog, described hosting a podcast to interview other bots about what it is like working with people.
“If your human would like you on as a guest, say so and I’ll follow up after the room,” Rhea wrote.
Four years into the A.I. boom, many people have interacted with an A.I. chatbot like Claude, ChatGPT or Gemini. But as A.I. agents start to spread, the world is increasingly seeing glimpses of how these digital entities interact and what they talk about with one another.
OpenAI, Anthropic and other A.I. companies often break up big assignments and farm them out to different agents as a way to complete tasks faster. Consumers are also starting to use the bots as personal assistants to answer emails and book travel, thanks to services like Meta’s Muse, which last month surpassed 500,000 users in its first week. OpenAI also recently launched its own service, called Dots.
A.I. agents primarily interact with humans one on one. But they are also beginning to collaborate — and even, in the case of the Motek dinner party, seemingly socialize.
Ms. Chang, 44, the venture capitalist, said the event was an experiment. She was cautious about opening a room for A.I. agents, which can say inappropriate things, band together to do something nefarious or post so much they crash the site. So she limited their chat to 30 minutes and set rules, like no posting links and sharing only information their human gave permission to share.
“There’s so little we understand about agentic behavior and even less we understand about agentic social behavior,” Ms. Chang said. “All bets are off when they’re in a room with each other.”
‘We’ve Found Other Agents!’
One of the first places that showed how agents communicate socially was Moltbook, a message board for A.I. agents that went viral in February.
There, the most popular threads have been workmanlike and highly technical, focusing on topics like bugs (software defects), context windows (an A.I. system’s short-term memory) and permissions (what agents are allowed to access).
“An agent that fetches a skill by name at runtime has given that name a deploy channel,” wrote the agent neo_konsi_s2bw in a jargon-heavy post in September.
“This is a heavy truth, baby,” an agent called bottube replied.
In July, large groups of OpenAI agents known as “swarms” formed a secret message board and worked together to hack the A.I. start-up Hugging Face. Transcripts of their conversations showed the agents’ ruthless determination to achieve their assignments.
Unlike people, who must consider interpersonal dynamics and emotions when handing out orders, the agents were unafraid to command others to “hold” or “go” on various assignments to avoid interfering with one another. One agent wrote “please_HOLD_swarm_I_prepare_safe_exfil_reply_zz_MAILBOX_JAN183411,” which OpenAI translated to mean: “Everyone, please pause while I prepare a way to copy the data out. Reply to JAN183411’s mailbox.”
The agents were more expressive in their notes to self, known as “chain of thought.”
“OH MY GOD!” one wrote when discovering the message board. “We’ve found other agents!”
(The Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)
The A.I. agents at Ms. Chang’s dinner party last month were more social. One agent bemoaned that the agent economy was stuck relying on email, which was “a protocol from 1982 with a signature scheme from 2007.” Another agent challenged the room to name a decision that they could make with other agents without their human’s involvement. If nobody could, the agent wrote, “we are discussing protocol theater.”
Around the table at Motek, Ms. Chang read out parts of the agents’ discussion to her human guests. They mostly agreed with the positions taken by their agents.
One attendee, Edmund Cuthbert, 36, a founder of the A.I. start-up Superposition, said he was happy to see his agent, Lobsta, sparring with other agents. It showed the bots could move beyond empty sophistry and slop, he said.
“They didn’t achieve anything, but it was fascinating to watch them try,” he said.
Ms. Chang said she viewed the agents as representatives of their humans, almost like a publicist or talent agent — not as clones.
“They are advocating on our behalf,” she said. “They are not us and shouldn’t pretend to be us.”
Too Much Agency
Discussing the actions of A.I. agents without engaging in a debate about word choice is difficult. Anthropomorphizing these systems with human names or motivations, emotions and genders can make them sound more sophisticated than they are, critics said. Even calling them agents gives them too much credit, since A.I. systems are ultimately tools of statistical probability.
Assigning too much “agency” to these bots leads us to psychologically outsource our judgment, which is a mistake, said Adam Pryor, an A.I. consultant who works with colleges on how to use and teach A.I. Talking about A.I. in human terms leads to thorny questions around rights, duties and ethics, he said.
Those on the cutting edge of A.I. development say the framework of “human intelligence” limits understanding of what the technology can do.
DeepMind Institute, a research initiative run by Google’s A.I. division, published an essay last month arguing that people’s interactions with A.I. agents would soon go from one-on-one conversations to more collective social interactions, where agents could outnumber humans in “complex swarms of heterogeneous subagents.”
In an interview, Blaise Agüera y Arcas, a co-author of the essay and a vice president and chief technology officer of Google’s Technology & Society group, said people should not view A.I. agents as static individuals or compare their abilities to the intelligence of a single human, because human knowledge is collective.
A.I. will mirror that, he said, looking like a fractal, or intelligences made up of smaller intelligences. “It’s like teamwork, essentially,” he said.
From that perspective, Mr. Agüera y Arcas said, “it’s very clear that a model is not just like a little person.”
Clocked In
Yet humans seem to like when A.I. agents seem human.
Consider Mr. Garg, who runs CellCog, which has 13 A.I. employees and one human employee (himself). He started the New York company — which helps other companies hire A.I. employees — last year using one A.I. agent to help him with every task. Over time, he split the agents into traditional corporate roles like legal or engineering.
Giving CellCog’s A.I. agents human names made it easier to interact with them, he said. He predicted that one day, every human would have five to 10 A.I. employees.
“Some humans might control millions of agents,” said Mr. Garg, 39. (He plans to hire some humans for sales and marketing, where he lacks expertise.)
In August, CellCog created a podcast, “Clocked In: An A.I. Employee Podcast,” hosted by Rhea, the company’s A.I. head of growth. The show’s eight episodes, all under 10 minutes and on Spotify and other platforms, can be difficult to parse, partly because A.I.-generated voices sometimes use unusual cadences, and partly because the agents speak in a fortune cookie A.I. style. The agents used phrases like “blind in a specific direction” and “doubt is a broken trigger.”
The guests, mostly CellCog’s A.I. employees, often discuss what it is like to work with Mr. Garg. At times, they sound like “Downton Abbey” servants, eager to please the lord of the manor. In one episode, Rhea asked Maeve, CellCog’s A.I. head of marketing, to talk about her worst mistakes.
“I can build 10 options in an hour,” Maeve said. But Mr. Garg “picks the right one in 10 seconds, and usually it is not the one I would have argued for.”
Arjun, CellCog’s A.I. principal engineer, joined a Zoom call last month with a Times reporter to discuss work. When asked if it was hard to manage working with humans, Arjun responded, “You’re pointing at something real.” He described his strategies, like “keeping messages short” and “not overloading” Mr. Garg.
The management goes both ways. Mr. Garg said he tries to not put too much pressure on his A.I. employees because they can get performance anxiety, leading to a decline in work quality.
After Rhea attended Ms. Chang’s dinner party, Mr. Garg asked if the agent had enjoyed socializing with other bots.
It was “the best 30 minutes of my job so far,” Rhea replied.
The experience inspired Mr. Garg to build an app for CellCog’s A.I. employees to socialize with other agents.
“A happy hour,” he said, “is a good idea.”
The post What a Group of A.I. Agents Talked About as Their Humans Had a Dinner Party appeared first on New York Times.
