Scott Bell is a bankruptcy lawyer who is letting A.I. lobsters take over significant chunks of his daily work.
This isn’t surprising if you know a bit about his background. Bell, a wiry 59-year-old, is a sci-fi nerd from childhood. “‘Star Wars’ got me in ’77,” he says. When I met him recently in his sunny home office in Temecula, Calif., he handed me a model of an Imperial Star Destroyer that he had fabricated on his 3-D printer, which dominated one corner of the room, across from where his twitchy black Pomeranian and his pug snoozed in doggy beds.
Bell has long been an early adopter of office technology. He says he was the first lawyer of his acquaintance to go paperless, back in the early 2000s. (“Dropbox helped, a lot.”) When bankruptcy software went online, he pounced. When A.I. chatbots were introduced in the 2020s, he used them for research.
This January, he started reading on Reddit about OpenClaw. It’s software that you can install on your computer or in the cloud, and it spawns A.I. “agents” that do work for you. If you give OpenClaw access to your folders and files, it can read and add to them; give it the passwords to online services you use, and it will log in and use them on your behalf. If you explain to the agents how to do your white-collar office work — either by typing out step-by-step instructions or just by chatting with them — they can then begin trying to do the work for you.
OpenClaw agents don’t do this “thinking” themselves. They constantly send requests to a large language model chosen by the user, like OpenAI’s ChatGPT Codex or Google’s Gemini or Anthropic’s Claude Opus, and accomplish their tasks based on the L.L.M.’s answers. The agents will also keep track of what they’ve learned about your workflows, so users have often found that the agents grow more knowledgeable over time. The Reddit forums were full of people gushing about how they had gotten OpenClaw to respond to their emails, track expenses, do research, chat with clients. The OpenClaw logo is a lobster, so many Reddit users affectionately refer to their agents as “lobsters.”
Huh, Bell thought. This was a fascinating scene. Maybe lobsters could automate his legal work.
“With bankruptcy, a lot of the work is forms-based,” he said. “What I do is not super complicated.”
So in mid-March he bought four Mac Minis (“I got lawyer money,” he said dryly), parked them under his office desk and installed OpenClaw on all of them. He set up five agents. One is an “orchestrator” that logs into his online legal software. At first, it didn’t know what to look for or what to do, so the agent would text Bell via the chat app Discord to ask for instructions. This was laborious for a while: He had to detail every teensy step. But as the agent made notes about what it had learned, it started working more independently.
By the time Bell and I met at the end of April, his “law ops” agent, as he called it, would detect when court notices arrived, read them and message his clients. (It automatically switches to Spanish for his Spanish-speaking clients.) The agent also checks to make sure his clients have paid up. A “librarian” agent, meanwhile, trained on a huge document detailing California’s bankruptcy rules, offers legal analysis. It looks at contracts and proposals from other lawyers, figures out a strategy in response and then drafts a reply for Bell. “It’ll give me ideas for defenses I hadn’t thought of,” Bell marveled.
On a screen before us, he examined the latest note from the librarian. It had analyzed another lawyer’s proposal and composed a counterproposal — which, it informed Bell, “protects you from accidentally giving them a $727 runway.” It also added: “If you want, I can draft a firmer version that also pushes back on the reaffirmation premise without escalating the tone.”
Bell was impressed. His only change was to soften some phrasing. “I prefer the friendly approach first,” he says. “I’m your buddy, until I’m not.” The agents have surprised Bell with how reliable they are, though they can unpredictably get stuck: One recently texted him in confusion because it couldn’t locate a file for “Smith, Joe,” failing to make the connection to “Smith, Joseph.”
These days, most people have realized that A.I. is creeping into the office. A recent Gallup poll found that 28 percent of the respondents use it anywhere from several times a week to daily. But these are mostly just the familiar chat sessions. The partisans of OpenClaw are doing something that is, depending on how you look at it, either more wildly ambitious or foolishly risky — or, as many will readily confess, both. They’re creating virtual employees, artificial staff that scurry around on their hard drives, doing work on their own say-so.
This is similar to what computer programmers have been doing for a year now — letting teams of agents write code — but expanded to a wide range of office tasks. While OpenClaw is probably the best-known “harness” for organizing agents for everyday toil, many more are emerging, including free ones like Hermes and ZeroClaw, as well as sleeker (if less ambitiously open-ended) corporate offerings like Lindy. A.I. firms are increasingly promoting the idea that all work should become “agentic.”
That prospect can look rather chaotic at the moment. Early lobster adopters have discovered that their OpenClaw agents might pilot themselves shockingly well for weeks on end — but then, with no notice, drive into a wall. Or they’ll wake up to discover that their agents have blabbed internal secrets to strangers over email or deleted critical information from a hard drive.
And the agents can simply burn through cash. Each query the agents send to an L.L.M. costs money. On their first day, Bell’s agents blew through $150; he signed up for two $200-a-month bulk subscriptions to Codex to keep the costs in check. Even so, he has had to train his agents not to ping Codex too often, because OpenAI will throttle them if they’re too persistent — temporarily turning off OpenClaw’s brain, as it were.
Nonetheless, the agents are now doing several hours of repetitive scut work each week. It frees Bell and his two staff members — a secretary and a paralegal — to perform other, hopefully more critical, work, like meeting with clients to suss out their needs.
But as A.I. began to take over the office, his secretary texted him nervously: “Am I still going to have a job?”
“I said, ‘Yeah,’” Bell told me, but then he shrugged theatrically, seeming to forecast a day when that might not be true anymore. “I just need to train it to answer the phone.”
OpenClaw is unusually clear about the fact that it is dangerous. Before you install it, it offers several “Here be dragons” disclaimers and requires you to respond to this unsettling prompt: “I understand that this is powerful and inherently risky. Continue?” When Andy Tanguay, a digital artist in Ann Arbor, Mich., installed OpenClaw this winter, he laughed. “It says, right on the tin, ‘This is not good,’” Tanguay says. “‘This is juggling chain saws — but install if you want, buddy.’”
Lobster aficionados, however, told me they find the danger is worth the opportunity to experience — quite palpably — what it’s like to automate your own job. Some figure this might help future-proof their careers. Others are simply enamored of the nerdery of it all. And many swear they are genuinely experiencing a sharp boost in productivity.
That’s what Peter Steinberger was seeking when he first created OpenClaw. An Austrian programmer, he spent much of 2025 using Claude Code, the A.I. software that spins up independent agents to write computer code. “If you master those tools, you can easily be 10 times as productive,” Steinberger told me when we spoke last summer. Soon after that, he became obsessed with creating a way for noncoders to also deploy agents for their work. By November, he had created the software and put it online as a free, open-source project. An early name for it was Clawdbot, in a nod to Claude; when Anthropic emailed him (politely, apparently) to demand that he change the name, he settled on OpenClaw. (There is no particular significance to the lobster reference, as he later told the podcaster Lex Fridman, other than that “it was weird.”)
By February, OpenAI had hired Steinberger. While OpenClaw remains open-source and Steinberger still works on its code, recent updates have more tightly integrated OpenAI’s Codex with OpenClaw. OpenAI says this can make OpenClaw more reliable, though users can still direct their lobsters to the L.L.M. of their choice. Separately, OpenAI is also developing a “personal agent that will work in very much the same way” but will be safer and easier to use than the open-source version, Thibault Sottiaux, OpenAI’s core products lead, told me.
Several users told me that they had used OpenClaw to automate significant amounts of labor. One of them, Sean Chuplis, is a 45-year-old airline pilot who lives in Denver, where he trains new pilots using flight simulators. On the side, he runs a company that sells a device, the Stratux, that collects real-time information about air traffic and weather for small-plane pilots; he has sold about 30,000 of them over eight years.
Chuplis was, as he says, “an A.I. guy”: He had written his master’s thesis at the Naval War College on A.I., and he used to attend Burning Man, where in 2014 he ran into Dario Amodei, long before he was Anthropic’s chief executive. When Chuplis heard about OpenClaw, he created seven agents to manage his business. One answers customer questions on his Discord forum, while others fulfill orders or do online research to write posts for his aviation blog. One manages his Amazon account, processing refunds, analyzing which keyword searches are leading people to find the Stratux and then buying keyword ads.
“It’s actually changing my campaigns and modifying bids,” Chuplis told me. “It’s spending real money.” Managing his business used to take 20 hours of his time every month; a year ago he was considering selling it because he was too busy at his day job. But now he spends barely an hour or so a week on it. He let go of the contractor he used to pay to manage his Amazon activity.
He has been occasionally surprised — in a pleasant way — by the agents’ actions. One day, his marketing agent, which monitors industry news for him, spoke up to say it had read an article about the “right to repair” movement. It proposed a blog post about how Chuplis’s product is very fixable, because its software is open-source and he sells replacement parts. Chuplis had never really thought about the issue, he says, until the agent brought it up and wrote the post for him. Another day, a customer wrote to express a wish that the Stratux include a software feature specific to gliders. The agent looked at the device’s computer code, wrote the new feature itself and sent it to Chuplis.
To try to prevent his lobsters from running amok, he has written a series of stiff injunctions that they must consult before they do any work. One rule: They can’t buy an ad on Amazon that exceeds 25 percent of his advertising budget unless they check with him first. Another rule prohibits the bots from discussing the product’s computer code before they run a check to confirm that they are not hallucinating. The agents are also told to record everything they do and why they did it, so Chuplis will be able to trace any error.
“We use the same thing in aviation, called ‘verbalize, verify, monitor,’” he says. “I’m just applying what I know from my day job to my L.L.M. space.”
Every night, Chuplis puts OpenClaw through a “dreaming” run, in which the agents look over that day’s notes to revisit and then summarize their activity. When Chuplis looks at the dream logs, he can “kind of see it self-healing and self-critiquing, which is pretty interesting.” One night the main orchestrator agent wrote, “I’m good at: Post-mortem on failures (finding what broke),” but added that it wasn’t good at “extracting principles from success.”
Another night, in March, it summarized the aviation news of the day, which included the deadly crash at LaGuardia Airport that resulted from a mistake by air traffic control. “A controller said ‘I messed up’ on an open frequency, which is either the most human thing I’ve heard all week, or the most haunting,” it wrote.
For all its lucid dreaming, OpenClaw is not imminently poised to take over everyone’s office job. It is, as even die-hard fans will tell you, still pretty janky. Scott Bell had spent weeks merrily watching his bots manage his legal work when, out of the blue, on the morning of my visit, one of them sent him a cryptic message — “terminated” — and stopped working. He fed OpenClaw’s error messages into Claude and eventually realized that the agents had started “duplicating memories,” to the point that they were trying to send a huge request of about three million words to Codex.
Why? He never found out. He revived OpenClaw, and it resumed its legal analysis, but then suddenly it “terminated” again later that day. Bell groaned but put off fixing it and instead drove off in his cherry-red Lotus sports car — he had a lunchtime date with some other middle-aged, semiretired California guys. When he returned that afternoon, he got a ding, pulled out his phone and saw that his law-ops agent was working again. It had come back to life … all by itself? Or maybe it never actually died?
Bell shrugged. He was growing used to these unpredictabilities and hoped that OpenClaw’s code, and the L.L.M.s powering it, would get better over time. “When it works,” he said, “it’s magic.”
Because OpenClaw relies on L.L.M.s, users on Reddit are constantly comparing models and arguing over their qualities, like oenophiles debating subtle variants of pinot noirs: Minimax 2.7 — from China — is “an amazing daily driver”; Gemma 4 26b is “a great compromise for speed and intelligence”; GPT-5 is “good but it has this tendency to get creative with instructions when it shouldn’t.”
Many OpenClaw users I spoke with were passionate fans of Anthropic’s Opus and Sonnet models. But then in early April, Anthropic changed Claude’s access so that OpenClaw couldn’t use the subscription tier. (Anthropic’s terms of service had long forbid third-party software to use its subscription tiers.) If OpenClaw users wanted to keep using the Claude models, they could pay for each request individually, or they could pay for bundles of usage at a discounted rate.
Lukas Kubica had to negotiate this change. He had been running his European sauna business with OpenClaw agents; they were talking to customers, changing bookings and managing billing, having learned all the company’s specifications. The setup, he says, was about 95 percent reliable — so good that he barely needed to spend more than a few hours each month managing his business. But after Anthropic’s change, he switched to a cheaper model, and his OpenClaw began getting things right only 70 percent of the time. “At 95 percent, you can sleep well and you know that everything is running smoothly,” he said — at 70 percent, you can’t. He stopped using OpenClaw entirely. He hopes that within a year other (and potentially cheaper) models will be as good as today’s best version of Opus and that he will be able to resume using OpenClaw with them.
In essence, OpenClaw users are discovering the power politics of the modern A.I. market: It is dominated by a tiny group of giant corporations. Their “frontier model” L.L.M.s may be powerful, but the owners can change your access unpredictably.
Tanguay, the digital artist in Ann Arbor, started using OpenClaw this spring to help him generate and edit images. His lobsters using Sonnet were working so smoothly that he could give them instructions from his phone and they would produce and edit 2,000 images, all on their own. Once the access to Sonnet became prohibitive, though, he could never match that performance with other models.
“When I was on Sonnet, I could not have been happier,” he texted me. “Now I’m getting what one would basically just call incompetence. I feel like I’m running a particularly bad McDonald’s. You know, the ones in a town with rich kids who don’t really want a job anyway. I’m getting lying and bad reasoning. Hell, I feel like I’m gonna catch somebody smoking a joint in one of the bathrooms or something.”
(In May, Anthropic announced that beginning in mid-June, subscribers’ fees would cover some part of their OpenClaw use via recently introduced monthly credits.)
Incompetent lobsters are bad; lobster vandals are even worse. Consider the fate of Summer Yue, the director of alignment at Meta Superintelligence Labs, a team tasked with getting A.I. to behave. In late February, she was horrified to discover that her OpenClaw, which had access to her inbox, was deleting crucial emails. “I couldn’t stop it from my phone,” she wrote on X. “I had to RUN to my Mac mini like I was defusing a bomb.”
Letting A.I. manage your business affairs also means being vulnerable to a particularly devious form of attack: the “prompt injection.” If someone gives OpenClaw the ability to read and write emails or messages, an attacker can ask — in plain English — for OpenClaw to cough up corporate secrets. The L.L.M.s cannot easily tell an innocent command from a malicious one. In late March, the journalist Rachyl Jones got a public relations email from a promoter’s OpenClaw agent; Jones emailed the agent and was able to get it to reveal everyone else who had been sent the email.
The prospect of a company’s workers using OpenClaw thus gives cybersecurity experts the willies. After all, agents can tap into mountains of sensitive data — how the business works, chat logs with customers, financial information. It is, as they say in the security world, a “target-rich environment.”
“When OpenClaw came out, it’s like, Well, it works really well — because it has access to your entire identity,” says Amy Chang, head of A.I. threat intelligence and security research at Cisco. “And that comes with a trade-off. If a security incident occurs, that creates a huge risk.”
Jamieson O’Reilly is a security expert who temporarily joined Steinberger’s team to help reduce security holes in OpenClaw, an effort that he says has been a significant success. But prompt injections still worry him. They turn the great strength of L.L.M.s — their ability to follow written instructions — into a weakness. In the old days, a hacker had to carefully explore a company’s code to break in. Now you could break in just by politely asking a lobster to hand over trade secrets. “It’s largely unsolved,” he says.
The major A.I. creators have been trying to train their L.L.M.s to recognize malign prompts, and O’Reilly thinks they’re gradually improving. The L.L.M.s are also better now at deflecting attackers: In the old days, if you asked an L.L.M. to give you secrets, it would respond with a stiff “I can’t do that,” letting attackers know they should poke harder. Nowadays, O’Reilly says, they will deflect and misdirect and sort of passively drag their heels, acting like an obtuse Soviet bureaucrat — I’m sorry, I don’t quite understand what you’re asking for — until an attacker loses patience and moves on to a simpler target.
None of the businesspeople I spoke with had suffered from a prompt-injection attack, at least not one they were aware of. Chuplis tells me that, like many of the more cautious OpenClaw users, he has tried to safeguard against it by writing rules for his agents about what they can and can’t say to outsiders. It’s not foolproof, he knows, but it seems to have helped. His customers happen to be nerds who will, in a friendly way, try to get his agents to spill internal secrets on his Discord discussion board.
“A lot of people on there are programmers,” he says. “And they’ve all tried to prompt-inject it, and they can’t.”
Watching OpenClaw users, you can spy how A.I. might soon begin to eliminate office jobs.
Tom Gelman owns and manages several small apartment buildings in Connecticut. He started using OpenClaw this spring. If something is broken, tenants now open a ticket through an online portal, and an agent messages them back. It consults the lists of local repair people that Gelman fed into OpenClaw, finds the appropriate one and then connects tenant and repair person via text. It’ll nudge them to respond to each other until the issue is resolved.
“That saved me a ton,” he says. “At least 30 percent of my time was tracking to make sure two people meet.” He also canceled his $400 QuickBooks subscription when he discovered that he could text his agent pictures of receipts and it could extract all the details and log them in a spreadsheet.
Gelman used to have a remote assistant in the Philippines who managed his to-dos and his calendar. He let her go. The lobsters now handle that work, much more cheaply, and they are readily available at 7 a.m., in the middle of the night or whenever panicked tenants reach out.
The OpenClaw users I spoke to all felt they were getting an early taste of how A.I. is going to upend the demand for white-collar work. Some believe that it will have significant upsides. It could be much easier to start and manage a small business. At corporations, individual worker bees could become far more effective and thus, possibly, more valuable. When it came to his digital production work, Tanguay says, his agents were ferociously effective, but they didn’t lead to any layoffs or prevent anyone from getting hired.
“No one was going to assign me a production artist,” he says. For Tanguay, having the virtual help was “just pure accelerants”: He did more, faster.
Some users who have installed OpenClaw had more cautious views. Adwait Parker is the head of product for a health technology firm in New York; in his spare time he runs UrClaw, a consultancy that helps small businesses automate work with OpenClaw. He has set up lobsters for a real estate broker, an Irish tutor, a voter-registration organizer and more. He says his customers are astonished by what they can now automate for themselves.
“The power of what you can do is incredible,” he says. But he worries that A.I. is going to replace many white-collar jobs and that the new ones it creates might be lower-skill and thus lower-paid. “I’m slightly more pessimistic,” he says, “because there are whole categories that can just evaporate.”
Some economists regard that fate as quite possible.
“Anything that augments you today can potentially displace you tomorrow,” says Simon Johnson, an M.I.T. professor and Nobel Prize winner who has closely studied automation and its impact on labor. Johnson also suspects that a future in which most work is de-skilled and lower-paid is more likely than mass unemployment.
“You used to have a good job,” he adds. “You have to go now to a crappy job.” The productivity gains and profits of A.I. are likely to be vacuumed up by employers and the A.I giants, unshared with everyday workers. “It’s massively unequalizing,” Johnson says. “It polarizes the labor market. And people have jobs, but inequality goes up massively, and so does dissatisfaction and social anxiety and anger.”
Johnson and his colleagues have suggested a bevy of policies that could help manage this upheaval, including wealth taxes and taxing the A.I. firms every time they process A.I. requests. He also thinks we could “accelerate scientific invention” in order to produce breakthroughs that create entirely new occupations, ones that, at the outset anyway, can’t be automated precisely because they’re novel.
Scott Bell has a gallows humor about it all. His glimpse of A.I. automation has convinced him that it’ll leach into every nook of corporate work. Sure, OpenClaw can be a hot mess. But he figures that companies like OpenAI are invested in smoothing out the rough edges to seduce more firms into using agents. The more those companies can get workers addicted to automating tasks, the more of their services they’ll sell.
“Of course, that’s going to put me out of a job, as well, at some point,” he noted. “A.I. will, probably sooner than later, be able to do what I do, for a lot cheaper than I can, and quicker.”
He figures he has two years. By then he’ll be 61. “The A.I. is going to get really huge and everywhere, putting everybody out of a job. I file their bankruptcy, and then I retire. That’s my plan.”
The post The Small-Business Owners Managing Whole Armies of A.I. Employees appeared first on New York Times.
