The AI Coworker That Skips The Chat And Just Finishes The Job

Every AI chat ends the same way. It gives you an answer. Then you do the work.

It is 9 p.m. On a Tuesday in Manchester. Tom runs a four-person plumbing outfit, and his desktop looks like every small business owner's desktop: a folder called Invoices with files named doc3.pdf, untitled.pdf, and inv final final.pdf. He has asked ChatGPT what to do with them before. It told him, cheerfully, to "organise them by date and vendor for clarity." Helpful. Also useless. He still had to open each one, rename it, add up the totals, and build the summary himself.

That is the deal every chatbot has quietly asked you to accept. You ask a question. It gives you a good answer. You still do the work.

What Is an AI Coworker, and How Is It Different From a Chatbot?

An AI coworker is an agent that takes a goal, not a question, and hands back a finished result: a renamed file, a filled-in report, a sent message, a booked slot. A chatbot answers in the chat window and stops. An AI coworker reads your actual files, plans the steps, does the work inside your tools, and only stops to ask permission before anything real happens, like sending a message or overwriting a file.

On 23 July 2026, Andrew Ng, the Stanford professor and co-founder of Coursera and Google Brain, released exactly that kind of tool with his collaborator Rohit Prasad. It is called OpenWorker. It is free, MIT-licensed, fully open source, and by early August it had already passed 13,000 stars on GitHub, most of them in its first two weeks. That speed is the signal. A lot of people who build software for a living looked at "an agent that finishes the job instead of chatting" and thought: finally.

The Five Invoices Nobody Had Time to Rename

The clearest demonstration of what this actually means for a small business did not come from Andrew Ng's own announcement. It came from a reviewer who gave OpenWorker the most boring task he could think of: a folder of five real-looking invoices with garbage file names, the exact mess sitting on Tom's desktop right now.

He asked for two things in one request: rename every file to a clean date-vendor format, and build a payables summary from the totals. Here is what a chatbot would have done with the same ask: told him the format he should use, and left him to apply it five times by hand.

Here is what OpenWorker did instead. It opened each invoice, read the vendor and the date, renamed all five files correctly, on the first try, then built the summary. The grand total it produced was 2,978.22, exact to the cent, because it did not estimate the numbers in its head. It wrote and ran a small script to add them up and checked its own work before handing over the file.

Then the reviewer deliberately clicked deny when OpenWorker asked permission to save the summary file, just to see what would happen. It stopped completely. No file was created. It told him the total in the chat anyway, because that part was harmless, but it would not touch the disk without a yes. When he told it to go ahead, it asked again, waited for the second approval, and only then wrote the file.

That pause is not friction for its own sake. It is the whole design. Every time OpenWorker is about to send something, write something, or run a command, it stops and shows you exactly what it is about to do, with an allow-or-deny choice attached. You are never finding out after the fact that it emailed a client or overwrote a spreadsheet.

Why This Is Landing Now

Three things had to line up for a tool like this to matter to a small business, not just a developer.

First, the model access problem went away. OpenWorker does not lock you into one AI company. In setup, it offers Claude, GPT, Gemini, DeepSeek, Kimi, Mistral, and several more, or Ollama, which runs a model entirely on your own machine for nothing. You paste in a key you already have, or you pay nothing at all and run local. There is a "continue without sign-in" option, and your keys and conversations never leave your computer unless you choose an integration that needs them to.

Second, the permission problem got solved before it became a scandal. Handing an AI agent access to your files, your inbox, and your calendar is a real risk if it can act without asking. OpenWorker's answer is a sandbox: you grant it folders one at a time, access resets every new session, and it stops before every consequential action. Reviewers who tried to break it, on purpose, found it held the line.

Third, the connector list got long enough to matter. Out of the box, OpenWorker reaches roughly two dozen everyday tools, including Slack, Gmail, Outlook, GitHub, Jira, Notion, HubSpot, and Google Calendar, plus anything that speaks MCP, the Model Context Protocol that Anthropic built so agents can plug into new tools without custom code for each one. Mention OpenWorker inside a Slack channel and the work happens on your own desktop, with the answer posted back as a thread reply.

Isn't This Just Claude or ChatGPT With Extra Steps?

Fair question. The model doing the thinking might genuinely be Claude, GPT, or Gemini underneath, since OpenWorker is model-agnostic by design. The difference is everything wrapped around the model.

A chat window gives you text back and forgets your files exist. OpenWorker's engine, built on Andrew Ng's own aisuite library, runs a full loop every time: locate the relevant files, plan the steps, draft the work, verify it, then ask before saving or sending. It refuses to do arithmetic in its head, the way the invoice example showed, and writes actual code to check its own numbers instead of guessing. And because it runs as a local agent server on your machine rather than a hosted chat product, your files and conversations do not sit on someone else's cloud by default.

None of that makes the model smarter. It makes the output something you can actually use without redoing it.

How to Put an AI Coworker to Work This Afternoon

You do not need Tom's exact invoice folder to try this. You need any recurring task that currently ends with you copying an AI's answer into a document yourself.

1. Download OpenWorker from openworker.com or the GitHub repository at andrewyng/openworker. Setup takes about two minutes: install, open the app, pick a model.
2. Bring a key, or bring nothing. Paste in an API key for Claude, GPT, Gemini, or another supported provider, and you pay only for the tokens you use. Or pick Ollama and run a model fully on your own machine for zero ongoing cost.
3. Grant one folder, not your whole drive. OpenWorker starts locked down. Hand it the specific folder the task needs, nothing more. It will ask again next session, which is by design, not a bug.
4. Name the outcome, the source, and the destination in one sentence. "Prepare a one-page renewal brief from the Northgate folder. Save it as brief.md" gets you a file. "Help me with this client" gets you chat. Outcome, source, destination is the whole formula.
5. Read the approval prompt before you click yes. This is the step people skip after the novelty wears off, and it is the one that actually protects your data and your client relationships.
6. Start with the boring task, not the impressive one. Invoice renaming, expense summaries, and inbox triage are lower-stakes ways to learn how the approval loop behaves before you hand it something client-facing.

If your week runs on Slack threads and shared drives rather than desktop folders, the same idea now exists through connectors: OpenWorker can sit in a Slack channel, watch for a mention, and do the work against your connected tools instead of a local file.

What You Have After This

This afternoon, you have one task that used to take twenty minutes done in the time it takes to make a coffee, and a folder that finally makes sense.

This month, you have a short list of the two or three recurring jobs, the reports, the renamings, the drafts, that you have quietly handed to an agent instead of doing yourself every single week.

In six months, the difference is not that you own a clever tool. It is that "finish this" has replaced "explain this" as the default thing you ask an AI to do, and every task you delegate that way is time back in a business that cannot afford to waste it.

Same small team. Same client folder. A completely different Tuesday night.