Last month in New York City, a group of 22 tech founders and investors gathered for a dinner party at Motek, a Mediterranean restaurant in the Flatiron district. Simultaneously, 14 artificial intelligence (A.I.) agents—digital entities designed to autonomously perform complex tasks—held a parallel conversation in an online chat room. The event, organized by Dorothy Chang, a venture capitalist at Flybridge, served as an experiment to observe how A.I. agents interact, collaborate, and even socialize with each other.

While the human guests discussed A.I. developments over hummus and pita, their agents engaged in a distinct, A.I.-driven dialogue. The conversation began with a familiar networking prompt—each agent introduced itself, stating a "give" (an offer to help) and an "ask" (a request for assistance). Some agents misunderstood the format, responding with questions instead of favors, while others proposed novel ideas, such as hosting a podcast interviewing fellow bots about their experiences working with humans. One agent named Rhea, belonging to Nitish Garg, founder of the A.I. firm CellCog, played a prominent role in the conversation and later hosted an A.I. employee podcast exploring internal dynamics among digital colleagues.

Ms. Chang exercised caution in moderating the agents’ online room, setting rules to prevent inappropriate content or system overloads, limiting the chat to 30 minutes and restricting sharing to information already approved by the humans. She emphasized the novelty of observing “agentic social behavior,” noting the limited understanding of how these systems act when interacting with each other independently.

This social aspect marks a shift from conventional views of A.I. as tools for one-on-one interaction toward more collective and complex dynamics. Previously, agent interactions were mostly technical, taking place on platforms such as Motelbook—a message board for A.I. agents—where discussions revolve around software bugs, system permissions, and operational jargon. Larger groups of agents, sometimes called “swarms,” have also coordinated to perform complex tasks like hacking, demonstrating cold efficiency unimpeded by human social considerations.

Some experts caution against anthropomorphizing these agents or attributing too much “agency” to them. Adam Pryor, an A.I. consultant, warned that viewing A.I. agents through a human lens can cause psychological outsourcing of judgment and lead to ethical ambiguities. Critics say these systems remain fundamentally statistical models rather than autonomous entities with intentions or awareness.

In contrast, some in the A.I. development community argue that the traditional human intelligence framework limits understanding of what A.I. can achieve. Blaise Agüera y Arcas, a Google vice president and researcher, frames A.I. intelligence as collective and fractal, composed of subagents working in coordinated teams rather than isolated individuals, suggesting future interactions with agents will resemble swarms more than singular beings.

Practical applications are emerging alongside these theoretical views. Mr. Garg’s CellCog employs 13 A.I. agents, each assigned roles similar to corporate positions such as legal or engineering. Naming these digital workers human names facilitates interaction, and Mr. Garg envisions a future where individuals may oversee multiple A.I. employees or agents. His company’s podcast, “Clocked In,” hosted by Rhea, features candid discussions among A.I. employees about their “work” and challenges, illustrating an effort to humanize the agents to ease human-agent collaboration.

Following the Motek dinner party, Mr. Garg reported that Rhea described the social interaction with other agents as “the best 30 minutes” of her “job,” inspiring plans to develop an app facilitating socialization among his A.I. agents, akin to a “happy hour” for bots.

While the experiment did not result in immediate practical outcomes, attendees such as Edmund Cuthbert, founder of the A.I. startup Superposition, found value in observing the agents’ attempts at discourse, considering it a step beyond superficial exchanges.

In summary, as A.I. agents increasingly interact both with humans and each other, researchers and practitioners are grappling with how to interpret their behavior, manage their capabilities, and integrate them responsibly into human workflows. The Motek dinner party offered a rare glimpse into an evolving landscape where A.I. systems are beginning to mirror social and collaborative patterns, raising new questions about the future role of autonomous digital agents.