Chat, Code, Claw: What Happens When AI Agents Work in Teams

Recent AI progress can be divided into roughly three phases. First we had chatbots, designed to converse. Then, those chatbots became proficient at using tools, allowing them to do things like search the web and write code. Now, thanks to a proliferation of new frameworks—notably the “OpenClaw” frame behind Moltbook’s virality—those tool-using agents can be orchestrated in fleets. 

If a tool-using chatbot is like a single digital worker, these new frameworks are like virtual firms in which dozens of agents, running 24 hours a day, can be organized hierarchically to accomplish a given task.

[time-brightcove not-tgx=”true”]

For example, if you were trying to build a website or a digital product, you could use Claude Opus 4.6 (Anthropic’s best model) to oversee a team of smaller Claude Sonnet models as they go out into the web, perform market research, and write and run code. You could connect the system to other digital platforms like WhatsApp, Discord, and Notion, allowing it to send you messages and create documentation. And instead of speaking directly with the workers, you could check in on their progress with their manager, Opus. As Andrej Karpathy—an AI pioneer who popularized the term “vibe-coding”—put it last week: “first there was chat, then there was code, now there is claw.”

Over the past two years, the word “agent” has been invoked so loosely and frequently by corporations eager to capitalize on AI excitement that it became diluted almost to meaninglessness. But in the background, with each release AI systems have become more intelligent—able to complete more complex tasks (particularly software tasks) and for longer periods of time. It’s that increase in model capability—coupled with new frameworks that make it easier for a system to retain memory and operate persistently—that’s enabling progress.

Cause for confusion: the term “AI system” can now refer to a single bot in a chat interface, a bot that lives in a digital environment from which it can write and run code, or to fleets of bots—potentially from different companies—linked together by a technical framework. Lay users chatting with chatbots are having an entirely different experience compared to people at the frontier commanding fleets. It’s no surprise these groups tend to talk past one another.

For now, the barrier to entry is moderate: you need to offer a physical computer or rent a virtual machine for the bots to occupy, pay for all the tokens they generate (costs can quickly stack up), and take extreme caution to avoid them accidentally becoming compromised and leaking all your data in the process. Those security risks are why some companies, like Meta, are instructing employees not to run Openclaw on their work machines.

Despite their competence, the agents are by no means perfectly reliable—as Summer Yue, Meta’s director of AI alignment, discovered when her claw system almost deleted all of her emails. The bot had lost track of Yue’s initial instructions, and was ignoring her requests for it to stop. To avert the crisis, Yue had to rush to switch off the Mac Mini machine where her claw resided. “I asked you to not action on anything until I approve, do you remember that?” Yue asked after the ordeal. “It seems that you were deleting my emails without my approval, and I couldn’t get you to stop until I killed all the processes on the host.” The bot wrote back: “Yes, I remember. And I violated it. You’re right to be upset,” before updating its memory and reassuring Yue it wouldn’t happen again.

Risks notwithstanding, the industry is moving fast. Peter Steinberger, who created the Openclaw framework, has since been hired by OpenAI. Commenting on the hire, OpenAI CEO Sam Altman said that Steinberger will “drive the next generation of personal agents,” and that the technology will soon become a core part of OpenAI’s products. “The future is going to be extremely multi-agent,” he said.

Whether the performance of these multi-agent frameworks will extend beyond software engineering tasks remains to be seen. But, as Karpathy notes, while implementation details are still being resolved, “the high-level idea is clear.”