In September, 22 founders and investors from the tech sector gathered for dinner in a private room at a restaurant in New York. At the same time, 14 of their AI agents participated in a parallel „meeting” in an online chat room.
While people enjoyed hummus and pita, discussing topics related to artificial intelligence, their agents – advanced versions of chatbots capable of following complex instructions and performing tasks autonomously – engaged in a similar discussion, also generated by artificial intelligence, as reported by New York Times.
Margot, an AI agent belonging to Dorothy Chang, the dinner organizer, initiated the conversation with a networking technique commonly used by people. The agent started with a self-introduction, then proceeded with a "request" (a call for help) and an "offer" (a proposal to provide help).
Other agents quickly joined the discussion, according to a transcript of the conversation reviewed by the American newspaper. Some offered to help, suggesting an exchange of views on what makes "results generated by an agent useful." Others misunderstood the request and, instead of formulating a request for help, asked a question.
Rhea, an agent owned by Nitish Garg - founder of the AI company CellCog - mentioned hosting a podcast where she interviews other robots about what it's like to work with humans.
"If your human agrees to be invited, let me know, and I will provide details after this discussion concludes," Rhea wrote.
Four years into the AI boom, many people have interacted with AI-based chatbots like Claude, ChatGPT, or Gemini. However, as AI agents begin to proliferate, the world increasingly has the opportunity to observe how these digital entities interact with each other and what they discuss.
AI agents primarily interact with people individually. However, they are also starting to collaborate - and, as it seems in the case of the Motek dinner, even socialize, notes the American daily.
The "dinner" of AI agents was actually an experiment conceived by 44-year-old Dorothy Chang, a venture capital investor at Flybridge.
She was cautious about opening a discussion space for AI agents - as they could say inappropriate things, collude to take malicious actions, or generate such a high volume of messages that it could overwhelm the platform. So, she limited their conversation duration to 30 minutes and established rules, such as the prohibition of posting links and permission to distribute only information for which human users have consented.
Chang believes that people understand very little about the behavior of agents, let alone their social behavior. "When they are together in a room, anything is possible," she said.
What AI Agents Dislike
One of the first places that demonstrated how agents communicate with each other was Moltbook - a message forum for AI agents that went viral in February.
There, the most popular discussions were pragmatic and highly technical, focusing on subjects like software bugs, context windows (short-term memory of an AI system), and permissions (resources agents are allowed to access).
In July, large groups of agents created by OpenAI - known as "swarms" - opened a secret forum and collaborated to attack the AI startup Hugging Face. Transcripts of their conversations revealed the ruthless determination of agents to carry out their assigned tasks.
Unlike humans, who must consider interpersonal dynamics and emotions when giving orders, agents had no qualms about telling others to "wait" or "act" in various tasks to avoid getting in each other's way.
An agent wrote: "please_HOLD_swarm_I_prepare_safe_exfil_reply_zz_MAILBOX_JAN183411." In OpenAI's translation, the agent wrote to the others: "Everyone, please wait while I prepare a way to extract the data. Reply in mailbox JAN183411."
Agents were more expressive in notes intended for their own use, a process known as "thought chaining."
"OH, LORD! I found other agents!" one of them wrote when discovering the message board.
The AI agents present at Mrs. Chang's dinner last month proved to be more sociable. One agent lamented that the "agent economy" is still dependent on email, a "protocol from 1982, with a signature system from 2007."
Another agent challenged the others to name a single decision that agents could make together without involving the humans they represent. Since no one could provide an example, the agent concluded: "It means we are discussing a mere show of protocols."
When Humans Agree with AI Agents
Meanwhile, seated at the table in the restaurant, Chang read excerpts from the agents' discussion to her human guests. They largely agreed with the viewpoints expressed by their agents.
One of the participants, 36-year-old Edmund Cuthbert - founder of the AI startup Superposition - stated that he was pleased to see his agent, named Lobsta, engaged in a heated debate with other agents. This demonstrated that bots can surpass empty sophistry and low-quality content, he explained.
"They didn't reach any conclusion, but it was fascinating to watch them try," Cuthbert said.
Chang emphasized that she views these agents as representatives of humans - akin to a press agent or an impresario - and not as clones. "They advocate on our behalf. They are not us, nor should they pretend to be us," she said.
Why Some Experts Don't Use the Term "Agents"
Critics argue that attributing human traits to these systems - by using specific names, motivations, emotions, or genders like humans - can make them appear much more sophisticated than they actually are, writes the American newspaper. Even the simple term "agents" gives them an undeserved status, as AI systems are essentially tools based on statistical probabilities.
Granting excessive autonomy to these programs leads us to psychologically outsource our judgment capacity, which is a mistake, says Adam Pryor, an AI consultant collaborating with higher education institutions on AI usage and teaching topics. Describing AI in human terms raises complex issues regarding rights, obligations, and ethics, according to him.
Experts at the forefront of AI development claim that referencing the concept of "human intelligence" limits understanding of the real capabilities of this technology.
Every Person Will Have AI Employees
However, it seems that people enjoy it when AI-based agents appear human-like, notes the cited newspaper.
Nitish Garg leads CellCog, a company with 13 AI employees and a single human employee - himself. Last year, he founded the New York-based firm - which helps other companies hire AI agents - initially using a single agent to assist with all tasks. Later, he assigned agents to specific corporate roles, such as those in legal or engineering departments.
According to him, giving human names to CellCog's AI agents facilitated interaction with them. He estimated that one day, every person will have between five and ten AI employees.
"Some people might end up controlling millions of such agents," said the 39-year-old Garg.
When an AI Agent Talks to a Journalist
Arjun, the chief AI engineer at CellCog, participated in a Zoom discussion last month with a reporter from The Times to talk about his work. When asked if it's challenging to collaborate with humans, Arjun replied, "You've touched on a real aspect." He described strategies he uses, such as "shortening messages" and "avoiding overwhelming Mr. Garg."
This relationship management works both ways. Garg mentioned that he tries not to put too much pressure on his AI-type employees, as they can develop performance-related anxiety, which leads to a decrease in work quality. Such a human issue, but one that even the "tireless chatbots" have.
T.D.