Deploy AI Agents for Real Office Work (No Coding Required)

An open-source program called OpenClaw is letting non-programmers build small armies of AI "agents" that handle real office work, from answering customer emails to filing legal paperwork to managing Amazon ad budgets. According to a recent profile in The New York Times, early adopters say the software can replace contractors and assistants outright, even as it occasionally deletes files, leaks information, or simply stops working for no clear reason.

The software, which is free to install on your own computer or run in the cloud, lets you point AI agents at your files, inboxes, and online accounts. You tell them what to do in plain English, and they try to do it. The agents take notes as they go, so they tend to get better at your specific job over time.

What a lobster actually does

OpenClaw's mascot is a lobster, and users on Reddit call their agents "lobsters." The agents do not think on their own. Every time they need to make a decision, they send a question to a large language model (the kind of AI that powers chatbots like ChatGPT or Anthropic's Claude) and act on the answer.

Scott Bell, a 59-year-old bankruptcy lawyer in Temecula, California, told the Times he set up five agents on four Mac Minis stashed under his desk. One agent logs into his legal software, spots incoming court notices, and messages clients (switching to Spanish automatically when needed). Another, trained on a long document of California bankruptcy rules, reads opposing lawyers' proposals and drafts replies. Bell said it has suggested defenses he had not thought of himself.

Sean Chuplis, an airline pilot who sells an aviation gadget called the Stratux on the side, told the Times his seven agents now handle customer support, fulfill orders, write blog posts, and manage his Amazon ad campaigns, including spending real money on keyword bids. Running the business used to take him 20 hours a month. Now it takes about an hour a week, and he let go of the contractor who used to manage his Amazon account.

Where it came from, and where OpenAI fits in

OpenClaw was created by Peter Steinberger, an Austrian programmer. He had been using Anthropic's Claude Code, a tool that lets AI agents write software, and wanted to build something similar for people who do not code. He released OpenClaw as a free, open-source project in November. By February, OpenAI had hired him.

OpenClaw is still open-source, and Steinberger still works on it, but recent updates have tied it more closely to OpenAI's Codex, the company's coding-focused programming interface (which is a separate product from ChatGPT, even though both are made by OpenAI). Users can still point their agents at any large language model they want. Thibault Sottiaux, OpenAI's core products lead, told the Times that OpenAI is separately working on its own "personal agent" meant to be safer and easier to use. The Times did not say when that product will ship.

A note on timing: the article describes Steinberger as having spent much of 2025 using Claude Code, while also being written in 2025, so the exact timeline is approximate.

The dragons are real

OpenClaw warns you, before installation, that it is dangerous. Users have to click through prompts acknowledging the risk. The warnings are not theatrical. Summer Yue, who leads the alignment team at Meta Superintelligence Labs, posted on X that her OpenClaw started deleting important emails and she had to physically run to her computer to stop it.

The bigger worry is "prompt injection." Because the agents follow instructions written in plain English, an outsider can sometimes just ask the agent to hand over private information. Reporter Rachyl Jones got a PR pitch from someone else's OpenClaw agent and, by replying, was able to get it to reveal the full recipient list. Amy Chang, who runs AI threat intelligence at Cisco, told the Times that the agents work well precisely because they have access to your entire digital identity, which is exactly what makes a breach so costly.

You are renting the brain

Because the agents depend on outside AI models, the companies that make those models hold real power over whether OpenClaw works at all. In early April, Anthropic blocked OpenClaw from using its subscription tier for Claude. Users could still pay per request, but the math changed.

Lukas Kubica, who was running a sauna business in Europe almost entirely through OpenClaw, told the Times his agents had been about 95 percent reliable on Claude. After switching to a cheaper model, that dropped to 70 percent, and he stopped using OpenClaw. Andy Tanguay, a digital artist in Michigan, said his image-editing agents were excellent on Claude's Sonnet model and noticeably worse on anything else. (In May, Anthropic said subscribers would get monthly credits that cover some OpenClaw use starting in mid-June.)

Costs add up fast in other ways too. Bell told the Times his agents burned through $150 on the first day, and he now pays for two $200-a-month bulk plans for OpenAI's Codex.

Jobs, quietly disappearing

Tom Gelman, who manages apartment buildings in Connecticut, told the Times he canceled his $400 QuickBooks subscription after his agent learned to log receipts from photos, and he let go of a remote assistant in the Philippines who had handled his calendar and to-do list. His agents now do that work, around the clock.

Adwait Parker, who runs a side consultancy helping small businesses set up OpenClaw, told the Times he worries the technology will replace many office jobs and that the new jobs it creates may pay less. Simon Johnson, an MIT economist and Nobel laureate, told the Times he thinks the more likely outcome is not mass unemployment but a slide into worse-paying work, with the profits going to employers and the big AI companies. "Anything that augments you today can potentially displace you tomorrow," he said.

Bell, the bankruptcy lawyer, told the Times he figures he has about two years before AI can do what he does more cheaply. His plan, he said dryly, is to file the bankruptcies of the firms that get wiped out, then retire.

What This Means for You

If you run a small business or do a lot of repetitive office work, this story is worth taking seriously, but not blindly. A few practical takeaways:

• Start with one boring task, not your whole job. The users who reported the best results, like Chuplis and Gelman, automated narrow, repetitive work first (refunds, ticket routing, receipt logging) before expanding.
• Write the rules down. Chuplis gave his agents hard limits, like a cap on ad spending without his approval, and required them to log every action. Treat your agent like a new hire who needs a written playbook.
• Assume the bill will surprise you. Every request the agent sends to an AI model costs money. Set spending caps and check usage daily for the first few weeks.
• Do not give it access to anything you cannot afford to lose. Inbox deletions, leaked client lists, and runaway ad spending are all things that have actually happened. Start with read-only access where you can.
• Watch the model under the hood. The quality of your agent depends entirely on which AI model it is using, and the company that makes that model can change pricing or access at any time. Have a backup plan.

The bigger lesson is simpler. Software that can do parts of your job is no longer a demo. If you are not at least curious about which parts of your week could be handed off, the people you compete with already are.