Story
When an AI agent (an AI that can take actions for me) works in a website or app for me, I want to watch what it does and take the controls the moment it goes wrong. Then I can fix 1 thing without starting over.
Hannah, a nurse in Bristol, renewing her parking permit on her laptop on a day off.
- 01 Setting
Hannah asks the agent to renew her residents' parking permit on the council site.
- 02 Trigger
The agent picks a car from last year's permit - the car she sold in May.
- 03 Without it
A spinner runs, then says Done. She finds the old car on the form only at the end.
- 04 With it
She sees "Choosing the car: WR11 KTE" in the live view, taps Take over and picks the new car.
- 05 Outcome
She taps Hand back. The agent carries on from her choice and stops at payment.
Live demo
This demo follows a script, so it plays the same every time - there is no real AI behind it. Change a setting to see the same moment play out differently.
What happens, step by step
| State | The person sees | The AI is | When | Screen reader |
|---|---|---|---|---|
| 1 | The ask. | Starts. | 0 s | - |
| 2 | The page the agent is using, and 1 line of what it is doing now. | Works and narrates. | while it works | “Working. Live view open” |
| 3 | Each choice in words, as it happens: "Choosing the car: WR11 KTE". | Works. | each step | - |
| 4 | "You're in control · the agent is paused". The agent's cursor is gone. | Paused. Does not touch the page. | on tap, at any moment | “You have control. The agent is paused” |
| 5 | The person changes what they need to. The log shows what they did. | Paused. | - | - |
| 6 | "Carrying on with YB24 LMN" - the agent says it saw the change. | Resumes from the person's state, not from the start. | on tap | “Agent resumed” |
| 7 | A past-tense result, stopped before payment. | Idle. | when done | “Done” |
On each device
- phone
- The narration sits in the thread and the live view is a card under it. Take over opens the page full screen. Long tasks often run elsewhere and send a notification when they need the person.
- tablet
- As on desktop.
- desktop
- The chat and narration on the left, the live page on the right. Take over and Stop sit under the live view, always visible.
When to use it
- The agent clicks and types in real websites or apps, where a wrong field is easy to miss.
- Some steps need the person's own hands - a sign-in, a password, a card number.
- People are new to the agent and want to see it work before they trust it.
- The work takes an hour or runs at night. Nobody will watch it - report back instead. Use Scheduled and background agents instead.
- Only 1 step needs a person, and it is known in advance. Stop and ask at that step. Use Approval gate instead.
- There is no page to watch, only text. Show the steps instead. Use Thinking and processing steps instead.
The parts of the pattern
- ALive viewThe real page the agent is using, updated as it works.
- BNarration1 line of what the agent is doing now, with the real value.
- CTake overPauses the agent and gives the person the controls. Always there.
- DStopEnds the task. Different from take over - nothing resumes.
- EWhere it isThe field the agent is working in, marked so the eye finds it.
Do and Avoid
✓ Filled in your address
● Choosing the car: WR11 KTE
Take overStop
✓ Filled in your address
● Choosing the car: WR11 KTE
Stop
✎ You chose YB24 LMN
● Carrying on with YB24 LMN
✎ You chose YB24 LMN
● Opening the permits page
Writing and accessibility
- Narrate in the present tense with the real value: "Choosing the car: WR11 KTE".
- Say who is in control in plain words: "You're in control · the agent is paused".
- After hand back, name what changed: "Carrying on with YB24 LMN".
- Keep "Take over" and "Stop" as 2 different words, because they do 2 different things.
- Put narration in a polite live region (the part of the page a screen reader reads out on its own, without interrupting). Announce only changes of step, not every keystroke.
- Announce control changes: "You have control. The agent is paused" and "Agent resumed".
- Take over and Stop are real buttons, reachable by keyboard at all times, not only on hover.
- Show who is in control with words and an icon, not only a coloured border.
Seen in real products
- ChatGPT agentWeb, apps and desktop · Jul 2025The agent narrates on screen. People can interrupt or take over the browser at any time, and Watch Mode is used for critical tasks such as email. OpenAI has since moved the agent into ChatGPT Work.OpenAI: Introducing ChatGPT agent (17 Jul 2025) ↗
- Gemini TaskWeb and apps · Nov 2026"View live" opens the browser Gemini is using. "Take over task" gives you the controls, and "Go back to Gemini" hands them back. Tasks pause at sign-in steps until you take over.Gemini Help: Tasks ↗
- Windows Copilot ActionsWindows · Nov 2025The agent works in a separate Agent Workspace desktop. People can check in on progress or take over the task at any time.Windows Insider Blog (17 Nov 2025) ↗
Each example comes from the product’s help pages, release notes or teardowns, with the month it was checked. Products change every month - tell us when 1 is out of date.
Check yourself
Seen this pattern done better, or a product that changed? Press Suggest next to any heading, or tell us about the whole page.