Chapter 15
What an agent is
The chapter in a minute
- An agent takes your goal, makes its own plan and acts on it with tools. A chat assistant answers, and an automation repeats fixed steps.
- Its work runs from the goal to a plan, then tools and actions, with a step log recording what it did.
- Agent features are new, change quickly and aren't offered to everyone, so this book leaves their names out.
- Give an agent short jobs you can check, such as finding, comparing and getting things ready. Keep paying, sending, sharing and signing in for yourself.
- A loose goal leaves an agent more to guess, and the less often it checks with you, the more it matters what it can reach.
Try it now
Be the agent
0 of 7 steps done
Step 1 of 7
Pick an errand you might hand to someone else. It could be booking a table for Saturday, finding the opening hours and price of your local pool, or comparing three ideas for a present. Keep it low stakes, and use a town or area, never your home address.
Tools and settings, as of a date
The book points to the ‘Agent options’ section of the companion page. Here it is in full, with the date each detail was checked. Open it on the companion page
Chapters 15 and 18 both close with this pointer. Chapter 18 adds that the section lists the current options "with the date each was checked". Ireland and the UK readers: not every feature is offered where you live.
Checked: 2 November 2026. This is the fastest-moving section on the page. Since the book's notes were taken, OpenAI has withdrawn its 2025 agent mode and replaced it with ChatGPT Work, and Anthropic's browser agent now runs inside Claude Cowork.
Current agent features, by provider
| Provider and feature | Plan, country and age limits | Page and date |
|---|---|---|
| OpenAI, ChatGPT Work (with its cloud browser) | Eligible paid plans, not Free or Go. OpenAI's launch note says Work "is available in all supported regions", and Ireland and the UK are both on its list of supported countries. Availability may vary while it rolls out. | ChatGPT Work and Codex (opens in a new tab), Using cloud browser in ChatGPT (opens in a new tab), ChatGPT release notes (opens in a new tab) (9 July 2026 entry), ChatGPT Supported Countries (opens in a new tab); all checked 2 November 2026 |
| OpenAI, dots (always-on agents that keep working between conversations) | Rolling out to eligible Pro and Business Premium users aged 18 or over. Pro access excludes the European Economic Area, Switzerland and the UK at launch, so not yet for personal accounts in Ireland or the UK. Created in the desktop app or on desktop web. | ChatGPT release notes (opens in a new tab) (29 September 2026 entry), checked 2 November 2026 |
| OpenAI, ChatGPT agent (the 2025 agent mode) | No longer available. The page now says to use ChatGPT Work instead. | ChatGPT agent (opens in a new tab), checked 2 November 2026 |
| Anthropic, Claude Cowork and Claude in Chrome | Paid plans (Pro, Max, Ultra, Team and Enterprise). The help pages give no country limits, and Ireland and the UK are on Anthropic's list of supported countries. | Get started with Claude Cowork (opens in a new tab), Get started with Claude in Chrome (opens in a new tab), Claude in Chrome permissions guide (opens in a new tab), supported countries (opens in a new tab); all checked 2 November 2026 |
| Google, tasks in Gemini | Age 18 or over; a personal Google Account; a Google AI Pro or Ultra subscription; Keep Activity on. Offered wherever the Gemini app is, except the European Economic Area, Nigeria, Switzerland and the UK. That rules out both Ireland (part of the European Economic Area) and the UK as of 2 November 2026. | Manage your tasks and workflows in Gemini Apps (opens in a new tab), checked 2 November 2026 |
How the options arrived, with dates
- 22 October 2024: developer "computer use" announced. Anthropic (opens in a new tab)
- 23 January 2025: first consumer browser agent, as a research preview in one country. Folded into the chat assistant on 17 July 2025. OpenAI (opens in a new tab)
- 17 July 2025: agent mode inside a chat assistant. OpenAI (opens in a new tab) (the page now carries a notice that the launch post is outdated)
- 25 August 2025: browser agent pilot. Anthropic (opens in a new tab)
- 18 November 2025: an agent in a second provider's assistant, in one country. Google (opens in a new tab)
- 9 July 2026: OpenAI launched ChatGPT Work on web and mobile for paid plans other than Free and Go, and later withdrew the 2025 agent mode. ChatGPT release notes (opens in a new tab) and ChatGPT agent (opens in a new tab)
- 29 September 2026: OpenAI began rolling out always-on agents, called dots, to some paid users; personal Pro access left out the European Economic Area, Switzerland and the UK at launch. ChatGPT release notes (opens in a new tab)
Which show a plan first, and what happens after a no
- Gemini: shows what it plans to do after you ask. Open the task and click the progress chip at the top to see each step marked completed, current or planned, and the files it read or changed. If you deny an action, it stops and does not try another way. Gemini tasks page (opens in a new tab)
- Claude Cowork and Claude in Chrome: Claude makes a plan from your request and shows its progress step by step. In Manual mode it asks before each action; in Automatically approve mode it checks each action for safety and blocks unsafe ones; in Skip all approvals mode nothing checks its actions. The side panel starts in Automatically approve mode, so choose Manual if you want to see each step first. Claude in Chrome permissions guide (opens in a new tab), Get started with Claude Cowork (opens in a new tab)
- ChatGPT Work: you can follow its progress, answer its questions, change direction and approve important actions. It asks you in the chat before anything hard to undo, such as confirming a booking or making a payment. ChatGPT Work and Codex (opens in a new tab), Using cloud browser in ChatGPT (opens in a new tab)
Stop and take-over controls
- Gemini: to cancel a task completely, click Stop in the Gemini text box. To stop only its web browsing, click Stop in the Gemini in Chrome side panel. On Android, the multi-step tasks page says you can "stop or take control of the task at any time". Gemini tasks page (opens in a new tab), multi-step tasks on Android (opens in a new tab)
- ChatGPT Work: OpenAI says to stop the task if it opens the wrong site or uses wrong information. If it gets stuck it asks you to take over its cloud browser, and you can ask to take over at any time. Using cloud browser in ChatGPT (opens in a new tab)
- Claude Cowork and Claude in Chrome: pause to ask you when something needs your approval, and you can step in mid-task to change direction. You can decline any single action. Get started with Claude Cowork (opens in a new tab)
Approval setting names and where they sit
See "Connections and permissions".
Sign-in handover and banking refusals
Every provider checked says the agent should stop and ask before it pays, sends in your name or signs in.
- Signing in: ChatGPT Work pauses and asks you to sign in through a secure form; OpenAI says the model can't see the username and password typed there, and you should never paste passwords, security codes or card details into the chat (cloud browser page (opens in a new tab)). Google's tasks page and OpenAI's earlier agent pages also hand sign-in over to you.
- Publishing or sharing personal data: Anthropic's pilot page says Claude asks users before high-risk actions like publishing, purchasing or sharing personal data.
- Refused or blocked outright by at least one provider: bank transfers and banking (OpenAI's 2025 agent pages); financial services sites (Anthropic's pilot); purchases and financial transactions, account creation, trades and permanent deletions in Claude in Chrome, whatever approval mode you choose (permissions guide (opens in a new tab), checked 2 November 2026).
Scheduling and recurring tasks
As of 2 November 2026, all three providers let tasks repeat:
- ChatGPT Work can run a task once, repeat it on a schedule or run it when something happens in a connected app, through Scheduled tasks. ChatGPT Work and Codex (opens in a new tab)
- Claude Cowork has a Scheduled list in its left sidebar, where you view and manage scheduled tasks. Get started with Claude Cowork (opens in a new tab)
- Google's Gemini tasks can repeat on a schedule. Google's page warns: "Avoid scheduling tasks you find sensitive", and if a schedule runs while you are offline, "you may not be able to stop Gemini from completing an unintended action". Gemini tasks page (opens in a new tab)
Makers' best-practice advice
- OpenAI: avoid vague instructions such as "Check my email and handle everything". ChatGPT agent (opens in a new tab), checked 2 November 2026.
- OpenAI: give clear instructions, check website addresses and confirmation requests, and use a connected app rather than the browser when an app can do the job. Using cloud browser in ChatGPT (opens in a new tab)
- Google: avoid scheduling sensitive tasks.
How to clear an agent task from an assistant's history
Chapter 18 suggests clearing a finished task if you'd like to. Here is how, as of 2 November 2026.
ChatGPT Work. A Work task is a chat, so you delete it the same way (Deleting and archiving chats in ChatGPT (opens in a new tab)):
- Open your chat history.
- Select the more options menu (•••) next to the task's chat.
- Select Delete, then confirm.
It leaves your account view at once and is deleted from OpenAI's systems within 30 days, with security and legal exceptions. Files saved separately in your Library are not deleted with it. If the task repeats, also remove it from Scheduled, where scheduled tasks are managed (ChatGPT Work and Codex (opens in a new tab)). To clear the sign-ins and cookies its cloud browser kept (Using cloud browser in ChatGPT (opens in a new tab)):
- Open Settings.
- Select Cloud browser.
- Under Browser data, select Cookies.
- Clear the data for all sites, or for one site.
Claude Cowork and Claude in Chrome. Side panel sessions in Chrome appear in your Cowork history like any other task (Get started with Claude in Chrome (opens in a new tab)). To delete one (Get started with Claude Cowork (opens in a new tab)):
- Find the task in your Tasks list.
- Click ⋮ next to it and choose Delete. Or tick one or more tasks in the list and click the bin icon.
The task leaves your history straight away and is deleted from Anthropic's storage within 30 days. To withdraw a website permission you gave Claude in Chrome, open the extension's settings and go to Permissions (permissions guide (opens in a new tab)).
Gemini tasks. To delete a task thread on a computer (Find and manage your tasks in Gemini Apps (opens in a new tab)):
- Go to gemini.google.com.
- In the sidebar, hover over the task.
- Click More, then Delete.
Deleting the thread removes the task, its Gemini Apps activity and any schedules attached to it. It doesn't remove files the task created or changed, or the data from Gemini's remote browser and remote code. Delete those separately (Gemini tasks page (opens in a new tab)): go to Settings, then Browser, and click Delete next to "Delete remote browser data"; then go to Settings, then Task settings, and click Delete under "Delete remote code execution".
Chapter 8 explains what deleting history does and doesn't reach.
Benchmark figures
Never in print. If you carry them here: see "Tool list" for the AI Index figures. Also Xue and colleagues, "An Illusion of Progress? Assessing the Current State of Web Agents" (COLM 2025): 300 tasks on 136 real websites, with average success down 31.6% from easy to medium tasks and lower again for hard tasks. arXiv 2504.01382 (opens in a new tab) Also, as the book notes, short tasks with few steps did best in that study.
Printable copies of Table 18.1 and Figure 18.3
For readers doing the chapter 18 walk-through on paper. These are the book's own material and are in the companion PDF. See "The companion PDF".
From the book
In the glossary wording, an agent is:An AI system that plans steps, uses tools and acts towards a goal with some independence. A chat assistant answers when asked, and an automation follows fixed steps; an agent chooses its own steps and takes actions.
Sources
- Building effective agents (opens in a new tab) · Anthropic, 2024
- Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku (opens in a new tab) · Anthropic, 2024
- Piloting Claude in Chrome (opens in a new tab) · Anthropic, 2025
- Claude in Chrome permissions guide (opens in a new tab) · Anthropic, 2026
- Get started with Claude in Chrome (opens in a new tab) · Anthropic, 2026
- Gemini 3 brings upgraded smarts and new capabilities to the Gemini app (opens in a new tab) · Google, 2025
- Manage your tasks & workflows in Gemini Apps (opens in a new tab) · Google, 2026
- The 2026 AI Index Report (opens in a new tab) · Stanford Institute for Human-Centered Artificial Intelligence, 2026
- ICO tech futures: Agentic AI (opens in a new tab) · Information Commissioner's Office, 2026
- Introducing ChatGPT agent: bridging research and action (opens in a new tab) · OpenAI, 2025
- Introducing Operator (opens in a new tab) · OpenAI, 2025
- ChatGPT agent (opens in a new tab) · OpenAI, 2026
- Toolformer: Language Models Can Teach Themselves to Use Tools (opens in a new tab) · Schick, Timo and Dwivedi-Yu, Jane and Dess\'i, Roberto and Raileanu, Roberta and Lomeli, Maria and Hambro, Eric and Zettlemoyer, Luke and Cancedda, Nicola and Scialom, Thomas, 2023
- An Illusion of Progress? Assessing the Current State of Web Agents (opens in a new tab) · Xue, Tianci and Qi, Weijian and Shi, Tianneng and Song, Chan Hee and Gou, Boyu and Song, Dawn and Sun, Huan and Su, Yu, 2025
- ReAct: Synergizing Reasoning and Acting in Language Models (opens in a new tab) · Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan, 2023