Computer use agent: ChatGPT in Chrome and the approval step

A computer use agent is an AI agent that operates software the way you do: it reads the page, clicks, types, fills in forms and sends messages inside apps where you are already signed in. On 15 October 2026 OpenAI put one into Chrome. ChatGPT in Work mode can now control your tabs. The question worth asking is no longer whether an agent can press the button. It is which buttons should wait for you.

What a computer use agent is

A chat assistant answers. A computer use agent acts. It does not need an API or a connector for each app, because it works the screen: the same tabs, forms and inboxes you would use, logged in as you.

That is the useful part and the risky part. An agent working in your signed-in Gmail, Salesforce or LinkedIn tab is acting with your identity. For the difference between talking to a model and handing it a task, see AI agent vs chatbot.

What changed on 15 October 2026

ChatGPT announced on X: "We're bringing the power of computer use to anyone who uses Chrome." In Chat, it can look across your open tabs and gather information. In Work, it "can control your Chrome tabs" for things "like sending messages, filling out forms, or downloading files." Public counts on that post, read on 16 October 2026, were 883,268 impressions and 6,634 likes. The extension is available on all ChatGPT plans.

OpenAI's browser extension documentation says ChatGPT "can read or act on sites where you're already signed in, such as LinkedIn, Salesforce, Gmail, or internal tools." Setup runs through Settings > Computer Use in the ChatGPT desktop app, and you pick the browser with an @-mention in a Work or Codex chat. Chrome, Edge, Brave, Opera and Vivaldi are listed.

Unite.AI reports that the extension's Chrome Web Store listing shows 7,000,000 users, and that the listing says ChatGPT will always ask for permission when referencing browser history and when downloading or uploading files. The Verge notes that this arrives after OpenAI's Atlas browser was shut down, and lists one of OpenAI's own sample prompts: "Use the details in this event brief to fill out the registration form, and stop before submitting."

That last instruction is the whole topic in eight words. OpenAI's example tells the agent where to stop.

Microsoft is heading the same way on the desktop. The Register reported on 7 October 2026 that Copilot's agentic features will get access to your entire file system, with a demo of Copilot finding a user's tax documents and drafting an email to an accountant with them attached. Microsoft says local agents will run in sandboxes with fine-grained permissions, and that Copilot will only search your documents if you ask.

Agents propose, humans approve

The same day as the Chrome update, Ledger launched Agent Intent. Ledger's announcement puts it in one line: "Agents propose. Humans approve. Hardware enforces." Agents keep working, every request for approval lands in one place, and you approve when you are ready. DegenerateNews' post on the launch carried the same line to a crypto audience.

The other side of the week was a warning. On 12 October 2026, Andi posted: "opus 5.5 just deleted my database". It was the dev database, "but it also had access to production." Public counts on that post, read on 16 October 2026, were 308,633 impressions and 1,125 likes. X's summary of the replies is AI Coding Agent Deletes Developer's Database, Prompts Safety Warnings.

Put those next to each other and the pattern is plain. The agent did what its access allowed. The approval step, not the model, is what decides whether a mistake reaches anyone.

What to approve when a computer use agent works for you

Reading is cheap to undo. Acting is not. A working rule for any computer use agent, ChatGPT in Chrome included:

• Anything sent to another person waits for you: messages, emails, comments, invites.
• Forms stop before submit. OpenAI's own sample prompt does this.
• Payments, purchases and anything that moves money need your confirmation, every time.
• Deletes and production systems are off limits unless a person signs off on the specific change.
• Uploads and downloads get a look, because that is where your files leave or enter the machine.

Everything else, such as comparing tabs, pulling details into a sheet or drafting a reply, can run without you. The output waits until you read it.

Where xSquads for Work fits

xSquads for Work gives each knowledge worker their own squad of specialised agents to delegate to and review, so they become 10x more productive. It is not a browser driver. The squad works on your own files and folders, and the work moves across a shared task board from Backlog to To Do to In Progress to Review to Done.

Review is the step this post is about. A card in Review is finished work waiting on you: read the change, approve it, or send it back with a note. You decide what lands, on your schedule.

The app is Mac now, with Windows coming soon. It is free via the waitlist. The roles, by job, are on squads. If you are comparing it with OpenAI's agent, ChatGPT agent alternative sets out what OpenAI documents and where xSquads differs.

When the agent's job is shipping code

A deleted database is what happens when an agent's access runs ahead of its review.

A managed software factory is a production line for code, operated for you. On this site that line is the code factory. Agents open pull requests you can read. A person reviews before code is merged. Nothing merges without your approval. You add or remove agents as the workload changes. The longer description is what a managed AI software factory does.

The change arrives as a diff you can read, not as something an agent already did to production.

Frequently Asked Questions

What is a computer use agent?

A computer use agent is an AI agent that operates software through the screen the way a person does. It reads pages, clicks, types, fills in forms and sends messages in apps where you are already signed in, without needing a separate API for each one.

Can ChatGPT control my Chrome tabs?

Yes. Since 15 October 2026, with OpenAI's extension installed, ChatGPT in Chat can look across your open tabs, and ChatGPT in Work can control them to send messages, fill out forms or download files. OpenAI says it is available on all ChatGPT plans.

Is a computer use agent safe on accounts I'm signed in to?

It acts with your sessions, so treat it like a colleague using your laptop. According to Unite.AI's reading of the Chrome Web Store listing, ChatGPT asks permission before referencing browser history and before downloading or uploading files. Keep sending, submitting, paying and deleting for yourself.

What is the difference between a computer use agent and a chatbot?

A chatbot answers in a conversation and you do the work. A computer use agent does the clicking and typing itself. That is why approval matters more for a computer use agent: its mistakes happen in your real accounts.

How does xSquads for Work handle approval?

Every finished card lands in Review on a shared task board. You read the change, then approve it or send it back with a note. The squad works on your own files and folders. The app is Mac now, with Windows coming soon, and free via the waitlist.

How is a managed software factory different from a computer use agent?

A computer use agent acts in your browser or desktop. A managed software factory is a production line for code, operated for you: agents open pull requests, a person reviews before code is merged, and nothing merges without your approval.

Related Reading

• AI agent workflow: a board and Review instead of ten chats.
• Personal AI agent: from one assistant to teammates you review.
• ChatGPT agent alternative and Microsoft Copilot alternatives.
• Managed software factory and the code factory.
• Squads and xSquads for Work.

Raina Rufus, Founder of xSquads

xSquads for Work is free. Join the waitlist and we'll send the download link.