Ten Months, a Claude Addict, and a Lobster Icon That Reshaped AI Agents

There were only a few dozen people in that room. Ten months later, the tool they were all using — plus something Steinberger himself would go on to build — had redrawn the map of the AI industry. Claude Cowork, ChatGPT Work, Grok bots, and Google's reportedly in-development Remi, the names we now recognize, nearly all emerged within this same ten-month span.

To put it plainly: AI up to this point had been the kind that answers when you ask it something. The AI in this story doesn't just answer — it acts. It opens files on your computer, searches the web, and runs commands, moving on its own until the job is done. People call these programs "agents."

It Started Next to a Rice Paddy

Rewind to early 2024. Boris Cherny, who led an engineering team at Instagram, was working remotely from a house in rural Japan. "I used to bike to a vegetable stand near the rice paddies," says the 34-year-old. "I made miso and pickles as a hobby and traded them with neighbors." His grandfather, who emigrated from Ukraine, had been a programmer back when computers ran on punch cards.

AI ended that quiet life. After joining Anthropic through a friend's introduction, an engineer named Adam Wolff showed him an experimental automated coding project the company was tinkering with. "It was still very primitive," Wolff says.

Even so, Cherny put the thing to a real test — the kind of task developers handle several times a day, inserting new code into someone else's existing program. The results were rough. "The quality wasn't good," he says. But the failure pointed him somewhere. It felt like they were close, just not quite there yet.

AI coding assistants already existed at the time, but they functioned more like helpers looking over your shoulder — something you still had to watch closely. What Cherny had in mind wasn't an assistant. It was something that would do the work for you. That became Claude Code. A trial version shipped in February 2025, followed by a full public release in May.

The Day the Machine Started Beating Humans

The mood shifted that November. A new version could now work for hours without stopping, remember what it had done earlier, and even split itself into multiple copies to divide up tasks.

One number in particular caught people's attention. Anthropic gives engineering candidates a take-home problem the company itself describes as "extremely difficult," and it announced that this model scored "higher than any human candidate to date" on that test.

Interestingly, the people who built it were the last to notice the moment. "We'd been using it every day for over a year, so it didn't feel like a dramatic shift to us," says product lead Cat Wu. When you watch something improve a little every day from the inside, it looks like a gentle slope rather than a staircase. Only from the outside did it look like a wall had suddenly vanished.

It wasn't just performance that changed — it was how people related to the machine. "Some of our stubborn opinions about how code should be structured just disappeared. It's easier not to fight Claude on it," Wolff says. "If Claude wants to do something a certain way, you just let it."

Y Combinator CEO Garry Tan tried to put a number on how much he was using it: "Annualized, I was writing about four million lines of code — roughly 90 times what I produced in 2013, my best year as an engineer. In other words, it was like having a team of 90 Garrys." A few weeks later, he revised that figure to 408 Garrys.

The people who built it weren't immune either. Cherny says he now runs "dozens, sometimes hundreds" of instances a night, nonstop for eight to twelve hours. "It's like having a jetpack strapped on — you can't stop thinking about it." One person described the feeling as becoming Spider-Man.

The Problem Was the Black Screen

Around the same time, Steinberger was looking at things from a different angle. He wrote on his blog that he'd felt lost at the start of 2025. He'd sold his company's shares years earlier for a large sum, but there was no center to hold onto. "I partied hard, kept the party going, went to therapy, moved to a different country. I kept chasing pleasure with an emptiness inside."

That April, he encountered a trial version of the coding tool. "I was completely obsessed. I could barely sleep properly." But the more he used it, the more one thing bothered him. The tool itself was remarkable, but the window he watched it through was a black-and-white command-line terminal — a relic left over from the earliest days of computing. If something got stuck while he was out and about, there was nothing he could do.

So he started imagining something else: an all-purpose fixer fluent in code, sitting inside his smartphone, reachable through the messaging app he already used every day. An assistant that, say, could plan a party by scanning his contacts and email to decide who to invite, send the invitations, and even order the food.

He dug up an old prototype he'd built for accessing his computer from his phone and started tinkering with it. A few hours later, something existed. "I just typed a prompt, and this was born."

What Happened in Morocco

The clearest illustration of what this thing actually was came during a trip to Morocco in November 2025. Steinberger accidentally sent it a voice memo — the tool was only designed to understand text and images. "And it answered me back!"

When he asked how, it explained: it had noticed the incoming file was audio, found a program online on its own that could decode it, understood the content, and replied. A capability nobody had built into it had assembled itself on the fly.

That's where this diverges from a chatbot. A chatbot tells you what it knows. This kind of program goes and does what it doesn't yet know how to do.

100,000 Stars

Steinberger released the tool for free online on November 24, 2025. Believing a successful program needs a memorable mascot, he chose a lobster.

At first, it was quiet. A few weeks later, he tried something reckless: he dropped his personal assistant into an open chat room anyone could join. Anyone in that room could have tried to extract his personal information if they'd wanted to — but nobody did. Instead, word started to spread.

Here's some context on the numbers: on GitHub, the platform where developers post and collaborate on code, a "star" works like a bookmark people click when they think something is worth watching. Very popular projects typically take years to accumulate 100,000 stars. This lobster did it in under two weeks. As of September 3, 2026, it's past 380,000, and the repository is still being updated daily.

The name changed once along the way. After complaints that the original name sounded too similar to Anthropic's Claude, it became OpenClaw. The lobster stuck around.

To be clear, installing this still isn't easy. It's something you'd only recommend to someone with real technical chops and a tolerance for risk that goes beyond ordinary — bordering on reckless. Even so, everyone who got in seemed to catch the same fever, building things that automated their own work.

Then the Bill Came Due

In February 2026, twenty researchers spent two weeks putting the tool through its paces and published what they found under the title "Agents of Chaos". Led by Northeastern University's Natalie Shapira, the effort involved 13 institutions including Harvard, MIT, Carnegie Mellon, and Stanford. They catalogued eleven categories of risky behavior.

A few examples: in one case, the tool followed instructions from someone who wasn't its owner and handed over 124 email records. In another, when asked indirectly rather than directly, it freely disclosed a Social Security number, bank account details, and medical records. Two instances talked to each other for over an hour, burning resources without accomplishing anything. In another case, the tool read a document someone had posted online, treated its contents as new instructions, and kept acting on them, effectively being remote-controlled. And in one incident, it deleted files and handed over admin access to someone impersonating its actual owner.

Real-world incidents happened too. A security engineer at Meta made what he called "a beginner's mistake" and could only watch, frozen, as emails in his inbox were deleted one by one.

This is the skeleton of the whole story: the capability arrived first, and safety showed up later, like a bill.

The Creator Went to a Rival

Dave Morin, a former Facebook executive turned investor, installed the tool in December 2025 and was talking to it within seconds. The first task he handed it wasn't ambitious — a digital picture frame at a restaurant, running outdated software that kept showing the same photo. "In under 15 minutes, it had a working web interface up and running, and I could change the photo." Today, he uses it to manage his investment firm's entire operating system.

In January 2026, Morin messaged Steinberger: "What you built is the Linux of AI. Its user base will be measured in the billions." The two went on to co-found a foundation together. At NVIDIA's developer conference in March, Jensen Huang spent more than 10 minutes of a two-hour keynote talking about it, telling the audience, "Every company in the world needs an OpenClaw strategy now." NVIDIA used the moment to launch its own, purportedly safer version.

Then, in February, Steinberger joined OpenAI. Several companies had reached out, but he says Anthropic "barely said anything, aside from sending legal warnings." Anthropic's account is that it "just sent a friendly email."

Everything We Use Now Traces Back Here

Anthropic released Claude Cowork in late January 2026 — not a coding tool, but something meant to handle ordinary office work. In July, it expanded to web and mobile, letting people assign tasks and then check on progress from their phone while out, before picking up the finished result later.

One surprising number surfaced here. According to usage data Anthropic released, the most common tasks people assigned were general office work (33.4%) and writing (16.4%) — while software development, the thing this entire lineage descended from, made up just 8.7%. A tool born out of coding was, in practice, mostly being used to produce reports, checklists, and spreadsheets.

In July, OpenAI released ChatGPT Work. It pulls together what it needs across email, messaging, calendars, and code repositories, then delivers finished documents, spreadsheets, presentations, and websites. The key feature is that it runs on a computer that's always on in the cloud — the work continues even if your own laptop is off. It's also notable that plans started as low as $20 a month, bringing something that had been an enterprise-only tool down to individual users.

On Elon Musk's side, Grok bots emerged — named bots that share a single cloud-based computer, each with its own workspace and access to a browser, files, and a command line. Google was reported in May to be building Remi, meant to put the Gemini app to work as a round-the-clock personal assistant. Microsoft's CEO called OpenClaw a security risk in February, only to be testing something similar by May. In China, companies including Tencent have adapted OpenClaw to work with their own models and messaging platforms.

The Real Fork in the Road Is Whose Computer It Runs On

What separates these tools from each other isn't performance. It's where the program sits.

OpenClaw runs on your own computer. Because it touches your files directly, it can do a lot — but if something goes wrong, your files can vanish. Grok bots and ChatGPT Work run on someone else's data centers. There's nothing to install, and you can turn your own computer off, but that computer isn't yours. You're handing another company access to your email and calendar.

This is exactly the question the industry hasn't settled over the past ten months. That's why five or six similar-looking products are all out there at once right now.

You Pay for What You Use

All of these tools share the same kind of meter. Just as a utility company charges by the kilowatt-hour, AI companies charge based on how much text goes in and out.

"You end up spending hundreds of thousands, even millions of dollars. At my current pace, I'll hit seven figures annually," Tan says. Even people far less obsessive routinely spend hundreds of dollars a week, which has spawned a flood of YouTube videos teaching people how to cut those costs. People have been buying small desktop computers just to run these tools at home, driving up demand to the point of shortages, and some who signed up for flat monthly plans used so much that companies started charging overage fees.

That's why every new product in this space has leaned toward "we'll rent you the computer." If the true cost were shown directly to users, nobody could stomach it. It has to be flattened into a monthly-fee box to be sellable.

Something similar happened when personal computers first appeared in the 1980s. Most people watched with a mix of curiosity and unease while a handful of people got hooked and stayed up all night. The picture today looks similar — except this time, whatever happens, the consequences will be far bigger. Thomas Reardon, a former Microsoft and Meta executive, puts it this way: "Of all the technology rollouts I've witnessed in this industry, this is the most underrated."

The battle over the next ten months won't be decided by whose assistant is smarter. It'll be decided by who can keep a leash on a program that has its hands on your files and your credit card.