AI Agents Plunged the Tech World into Chaos

It was August 2025 and Peter Steinberger was addressing a meetup in London called Claude Code Anonymous. Steinberger and some fellow addicts had arranged the event to network with people like themselves—techies swept up by coding tools such as Anthropic’s paradigm-busting Claude Code. “I dedicate pretty much all my waking time to this, yet it doesn’t feel enough,” he told the gathering in a cozy, brick-walled room. A few months later, Anthropic released a new version of Claude Code, and the ranks of Claudeholics exploded. Called Opus 4.5, it could handle more complicated programming tasks, retain much more in its memory, run for many hours on end, and manage a team of AI subagents. Anthropic has what it describes as a “notoriously difficult” take-home exam for prospective engineering hires; in a head-to-head comparison of those people and its models, Anthropic claimed that Opus 4.5 “scored higher than any human candidate ever,” which “raises questions on how AI will change engineering as a profession.” Countless coders spent the holidays in basements and dens, madly trying out this new toy that let them build software as if they’d unleashed a hundred clones. Or unlocked superpowers. “It feels like becoming Spider-Man,” one told me. For the 39-year-old Steinberger, who split his time between homes in London and Vienna, even this was not enough. In November 2025, he launched a tool that’s now called OpenClaw, a simple way to conjure a personal AI agent that exploits the advances of Claude Code or other coding tools. Give it access to your data, your apps, and maybe even your credit card, and it scours your cloud and ventures onto the web to do your bidding. It can run autonomously in the background and overcome obstacles with the persistence of the Terminator. Steinberger’s project took off midwinter. One indicator of popularity is the number of “stars” a code repository gets on Github. In less than two weeks, as users downloaded it and began feverishly building, the project racked up more than 100,000 stars. (As of early May, it stood at 366,000 stars.) With those two breakthroughs—the commercial product Claude Code and the open source OpenClaw—the long-awaited age of AI agents has suddenly arrived. At least for those technically proficient enough and perhaps foolhardy enough to go all-in on a messy, imperfect, and risky adventure. More than one Claudeholic tells me they feel they are living in the future. “AGI is here!” one fanatic told me, paraphrasing William Gibson’s famous quote. “It’s just not evenly distributed.” Back in the 1980s computer revolution, the general public tended to regard the new machines with a mix of curiosity and angst while the hackers were joyfully building. There’s a similar dynamic today, possibly with even more at stake. “It’s hard to explain how much of a sea change this is,” says Thomas Reardon, a former executive at Microsoft and Meta who now heads a startup focused on a different area of AI. “It’s the most underrated, massive release I’ve experienced in technology.” Soon we’ll all be experiencing it. On a recent podcast, Marc Andreessen, the guy who co-invented the browser and has cast himself as the ultimate techno-optimist and MAGA fan, made a proclamation that reflects Silicon Valley’s thinking: “It’s almost inevitable that this is the way people are going to use computers.” Left unsaid: It won’t be a choice. Soon after Cherny joined the company, an engineer named Adam Wolff showed him Anthropic’s work on automated coding. “It was very primitive,” Wolff says. But Cherny used the coding tool to do a pull request, a common activity in software engineering that attempts to merge new code into an existing code base. “It wasn’t a good PR,” says Wolff. But the attempt meant that good pull requests were possible—and that higher-level coding tasks might soon be automated. Cherny set out to make that a reality. OpenAI and Microsoft had been trumpeting their coding products since 2021, when OpenAI launched its first iteration of Codex. While those tools made programmers more productive, the products were limited and required careful supervision. Cherny envisioned an upgrade where the model understood the architecture of programming and had the wherewithal to solve problems. What Cherny and his team built became Claude Code. Anthropic released a preview in February 2025 and launched it in May 2025, with updates over the next months. For many people, however, the turning point came in November, when the company released Opus 4.5: It ran longer, was better at problem-solving, and could run whole teams of subagents, each working on a different part of a program. At first, the Claude Code team saw Opus 4.5 as an incremental improvement. “We’d been daily users for over a year, so it was less night and day for us,” says Cat Wu, Claude Code’s head of product. But Anthropic’s tool had reached the coding equivalent of escape velocity. While far from perfect—only fools deployed its work without vetting for errors—it now rivaled or outperformed what a human might come up with. “Some opinions we had about how to structure code have melted away because it’s easier not to fight Claude,” Wolff says. “If Claude wants to do something a certain way, you just let Claude do it.” Much like how OpenAI underestimated ChatGPT’s impact, Anthropic didn’t anticipate how the November release would galvanize the techie community. Garry Tan, CEO of the startup incubator Y Combinator, was a convert. “I went all the way to the edge of how you could get the most out of Claude Code,” he says. “I was coding at a rate of about 4 million lines of code a year, which is about 90X my best output as an engineer in 2013—basically a team of 90 Garrys.” (A few weeks after we spoke, he updated his estimate; he now thinks he’s up to 408 Garrys.) Ryan Petersen, CEO of the shipping and freight logistics company Flexport, found himself spending less time on C-suite activities—or his family—and more time playing with Claude Code. “There’s something about watching the agent just doing the work that is mind-blowing,” he says. As I spoke to Petersen, I got the impression that the global supply-chain crisis in the Strait of Hormuz was less a corporate emergency and more an unwelcome distraction from his sessions with Claude. “It’s sad, because I just want to spend all day building tech and applying AI and partnering with the teams,” he says. In a twist that Mary Shelley might appreciate, Cherny himself turned into a Claudeholic. “Most nights, I have dozens, sometimes hundreds, of agents running eight and 12 hours at a time,” he told me. “I have some agents that run for many days at a time, and they do things like rewrite the code base or improve the efficiency of code.” Or as he said to me at another point: “It’s like I have a jet pack. I can’t stop thinking about it.” “I was really addicted,” Steinberger says. “I had trouble sleeping.” As transformational as Claude Code was, it still required monitoring from the command line in a terminal, a persistent relic of computing’s Jurassic roots. If Claude hit a snag while Steinberger was out and about, he couldn’t immediately fix the problem. Steinberger started to imagine a code-savvy factotum with access to Claude or Codex that you could use on your phone, perhaps through Slack or WhatsApp. He figured that a digital assistant with access to your apps and data could automate a huge range of tasks. To plan a party, for example, it could look at your contacts and email, figure out whom to include, send invites, and order the food. Steinberger dusted off a prototype of a tool that would let you access your computer terminal on your phone, and started building using Codex. (-OpenAI had upgraded its own programming tool to compete with Claude Code, and Steinberger often preferred it.) After a few hours of tinkering, he had an agent that could make use of the best AI coding tools. “I just prompted it into existence,” he says. Steinberger was startled by the skill set of his creation. On a trip to Morocco in November 2025, he accidentally asked his agent a question using a voice memo. He had designed the agent to accept only text or images. “It just replied!” he says. “I asked, ‘How the hell did you do that?’” The agent responded that it recognized the input as an audio file and found programs to decode it, understand it, and act on it. “That was a moment where I was like, holy hell!” he says. He dubbed his tool Clawd, and in late November he released it on GitHub as open source. Since every great open source product has a mascot, he made Clawd’s a lobster. At first, the uptake was slow. But a few weeks in, he took a chance and introduced his agent to a public Discord. Anyone on the Discord could have used it to mine Steinberger’s personal data, but they didn’t; Clawd went viral, becoming the most popular open source project in Github’s history. Viewed on a chart, Clawd’s trajectory looks like a rocket launch—one stark vertical line. As interest rose, Anthropic decided that the name was too close to its own product, which eventually led Steinberger to rename Clawd to OpenClaw. The lobster mascot remains. Make no mistake: OpenClaw is not necessarily easy to set up. If you don’t have a modicum of technical chops and an unhealthy tolerance for risks, it’s not for you. But those who enter OpenClaw’s world—by pasting in the single line of code that starts the installation process—have gone bananas building artisanal services to automate their professional and personal tasks. If your work involves, say, placing a lot of orders and awaiting their delivery, you can now set up a system that finds the orders in your email, regularly pings FedEx or UPS for updates, and displays the status of all deliveries in a dashboard. When a package arrives, OpenClaw messages you. In the gleeful fog of creation, many OpenClaw fanatics blew past concerns about the risks of exposing their data to an ambitious robot. In a February paper, 20 AI researchers tested OpenClaw and found that it is, to cite the paper’s title, an agent of chaos. “Observed behaviors include unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions,” and numerous other alarming behaviors. Problems emerged in the wild as well. One Meta safety and security engineer made a “rookie mistake” in an OpenClaw project and watched in horror as her inbox began deleting all her mail. But for all its risks, OpenClaw helped expose a broad swath of techies to the power of an agent. With OpenClaw, the interface is your chat app—WhatsApp, Telegram, or iMessage, for example. “It made the models so much more accessible and nice to use, because suddenly, like, you hatch your agent. You give it a name. It’s much more relatable,” Steinberger says. Dave Morin, a former Facebook executive and now a VC, says OpenClaw was life-changing for him. Within seconds of installing it in December 2025, he was in conversation with his OpenClaw agent. The agent asked him a few personal questions and suggested, because he lives in Marin County and is involved with the woo-woo Esalen Institute, that its name should be Watts, after the late Sausalito-based hippie philosopher Alan Watts. Morin had a job for Watts. In his dining room, Morin had a set of digital photo frames with outdated software, and he was stuck looking at the same pictures. He asked OpenClaw to take a stab. “Within 15 minutes, I had a fully functioning web interface for the frames, and I was updating the photos,” he says. Morin has since used OpenClaw to manage the software that runs his VC company. As I talk to him, I get the feeling that Watts may be Morin’s best friend. Morin sent Steinberger a DM on January 11. “I’m in love with what you’ve built,” he wrote, gushing that Steinberger’s idea is bigger than the software. Morin and Steinberger began a friendship that became a collaborati

[truncated for AI cost control]