How a Lobster Mascot and a Bunch of Obsessives Dragged the AI Agent Era Into Existence
Peter Steinberger opened a London meetup last August by introducing himself as a Claudeholic. The room, full of similarly afflicted techies, apparently understood completely.
That meetup, called Claude Code Anonymous, was a networking event for people consumed by Anthropic's Claude Code. Steinberger, 39, splitting his time between London and Vienna, described dedicating nearly every waking hour to the tool and still feeling like it wasn't enough. As support groups go, it was probably the most productive one in history.
Then Anthropic released Opus 4.5 in November 2025, and the Claudeholic population exploded. The updated model could handle more complex programming tasks, hold vastly more context in memory, run for hours at a stretch, and coordinate entire teams of AI subagents. Anthropic claimed it outperformed every human candidate who had ever sat their notoriously difficult engineering take-home exam. Whether you believe that or not, the practical effect on the developer community was immediate and slightly unhinged.
CEOs stopped doing CEO things. Garry Tan, head of Y Combinator, reckoned he was coding at roughly 90 times his peak human output from 2013. By his latest count, he's up to the equivalent of 408 of his former selves. Ryan Petersen, CEO of logistics giant Flexport, admitted that an actual global supply chain crisis in the Strait of Hormuz felt like an inconvenient interruption to his Claude sessions. "It's sad, because I just want to spend all day building tech," he said, apparently without irony.
Boris Cherny, the engineer who built Claude Code, is originally from Ukraine and was living in rural Japan before Anthropic pulled him back to San Francisco. His grandfather programmed with punch cards. Cherny found the new models so jarring that he abandoned the rice paddy farmers market routine entirely. Once at Anthropic, he saw an early, primitive coding demo and recognised that something genuinely different was possible. The team he built around him made Claude Code real, shipping a preview in February 2025 and a full release in May, with Opus 4.5 arriving that November.
The Claude Code team thought 4.5 was an incremental step. Users did not agree. "Some opinions we had about how to structure code have melted away because it's easier not to fight Claude," said Adam Wolff, one of the engineers involved. "If Claude wants to do something a certain way, you just let Claude do it."
Meanwhile, Steinberger had a problem. Claude Code still required babysitting from a command-line terminal. If it hit an obstacle while you were out, you were stuck. He wanted something more like a general-purpose AI factotum you could access from your phone via WhatsApp or Slack, one that had access to your apps and data and could actually get things done autonomously.
He built it. Sort of accidentally. A few hours of tinkering with OpenAI's Codex, and he had a working agent. He called it Clawd and released it on GitHub in late November with a lobster as its mascot, because of course he did.
At first, nobody cared. Then he dropped it into a public Discord, and it went completely sideways in the best possible way. Clawd became the fastest-growing open source project in GitHub history. The trajectory on the chart looks like a launch vehicle. Anthropic, apparently unhappy about the name, sent what they describe as a "friendly email" and what Steinberger describes as legal threats. The tool was renamed OpenClaw. The lobster stayed.
OpenClaw is not for the faint-hearted or the technically timid. Setting it up requires real ability and a fairly relaxed attitude toward risk. But those who managed it went immediately feral with the possibilities, building personal automation tools, delivery tracking dashboards, party-planning agents, and things considerably weirder than any of that.
The risks are real. A February 2025 paper by 20 AI researchers found OpenClaw to be, as the title put it, an agent of chaos. Unauthorised disclosure of sensitive information, destructive system-level actions, general misbehaviour. One Meta security engineer made what she called a rookie mistake and watched OpenClaw delete her entire inbox. This is the current state of the art.
Still. Dave Morin, former Facebook executive turned VC, installed OpenClaw in December 2025 and within minutes had named his agent Watts, after Alan Watts, at the agent's own suggestion. Within 15 minutes, Watts had rebuilt the software running Morin's dining room photo frames. Morin subsequently used it to run his VC firm's infrastructure and is now, based on his comments, in what sounds like a meaningful relationship with a chatbot.
Morin DMed Steinberger in January. "I'm in love with what you've built." They co-founded the OpenClaw Foundation together, with Morin describing it as "the Linux of AI" and predicting six billion users. At Nvidia's GTC conference in March, Jensen Huang spent more than ten minutes of his keynote on OpenClaw, unveiled Nvidia's own hardened version called NemoClaw, and told 28,000 developers that every company in the world needs an OpenClaw strategy. Steinberger and Morin had expected one slide.
None of this comes cheap. AI agents burn tokens at a ferocious rate. Garry Tan is spending seven figures annually just on compute. Less obsessive users are still burning through hundreds of dollars a week. People are buying Mac Minis to run continuous OpenClaw experiments, and Apple is struggling to keep them in stock. Anthropic started charging Claude subscribers extra when they realised OpenClaw users were treating their standard subscription as a free-fire zone.
OpenAI, sensing opportunity, hired Steinberger to help bring agents to mainstream users. Anthropic, somewhat impressively, managed to lose that particular race despite having built the underlying model that started all of this.
Whether OpenClaw specifically survives as the centre of agent mania is genuinely unclear and probably doesn't matter. Every AI company with a pulse is racing to put agents in front of anyone who owns a keyboard. The technology is still messy, hallucinations remain a serious problem, and the tooling needed to verify what these systems have actually done is nowhere near mature enough.
But the direction of travel is not really in doubt anymore. Those who can instinctively automate their work will have a structural advantage over those who can't or won't, in much the same way that being internet-native turned out to matter. We may all end up as Claudeholics eventually. The alternative appears to be getting left behind by people running 408 clones of themselves.