The short answer
To use an AI agent, describe the finished outcome, give it the files or apps it needs, set what it may do without asking, then review its work before it leaves your hands. In November 2026 the main options are Claude, where Cowork now runs in every conversation on paid plans, ChatGPT Work and dots from OpenAI, Tasks in Gemini for Google AI Pro and Ultra, and Perplexity Computer, which runs on credits. Start with one small, recurring task.
What an AI agent is, in plain English
A chatbot answers. An AI agent does: you describe an outcome, and it plans the steps, uses tools such as a browser, your files or connected apps, checks its own progress and hands back finished work. Since September 2026 the big assistants have all moved this way. On 16 September 2026 Anthropic began merging chat with Claude Cowork, so any Claude conversation can take on a whole task. OpenAI has ChatGPT Work and always-on agents called dots. Google is rolling out Tasks in Gemini, and Perplexity has Perplexity Computer. The skill is no longer writing the perfect question; it is delegating well.
What you need first
- A task with a clear finished state. "A spreadsheet comparing these five suppliers with sources" is agent work. "Help me think about suppliers" is a chat.
- The material and access the task needs: a folder, a connected drive, a list of websites.
- The right plan. Most agent features need a paid plan, listed below.
- A rule for what the agent may do without asking, especially sending, deleting and paying.
The main AI agents you can use in November 2026
| Agent | Who gets it | Good for |
|---|---|---|
| Claude (with Cowork in every conversation) | Rolling out to Pro, Max and Ultra first; Team in beta; Enterprise where the Chat and Cowork unified setting is on | Reports, spreadsheets and decks from your files; scheduled and cloud tasks |
| ChatGPT Work | Free on desktop with limits; Plus on desktop, web and mobile | Documents, presentations and spreadsheets using connected apps and a browser |
| ChatGPT dots | Eligible Pro and Business Premium users aged 18+; Pro access excluded the EEA, Switzerland and the UK at launch | Always-on agents with their own cloud computer |
| Tasks in Gemini | Rolling out to Google AI Pro and Ultra; early access on Ultra in select countries | Work that continues across Google apps while you are away |
| Perplexity Computer | Paid plans, using credits | Multi-step research, documents and apps in a cloud sandbox |
| Copilot Cowork | Personal accounts: Microsoft 365 Premium or Pro (preview). Work accounts: a Microsoft 365 Copilot licence, billed by usage, once an admin turns it on | Delegating multi-step work in Microsoft 365 |
How to use an AI agent in 7 steps
- Describe the outcome, not the steps. Name the finished thing, the material to use and how you will judge it. Anthropic's help centre says you no longer need to break work into steps or pick a mode in Claude.
- Give it the material. In Claude Desktop you can give Claude access to a folder; granted folders are listed under Trusted folders. Connectors for other apps live under Customize, then Connectors.
- Set how much it may do alone. In Claude, a permission setting in the message box controls this: Auto is the default on Pro, Max, Ultra and Team, and Manual, which asks before actions, is the default on Enterprise. Start with Manual for sensitive folders. By default Claude asks before permanently deleting files.
- Choose the effort. In Perplexity Computer, Effort Mode runs from Light through Standard and High to Ultra; higher effort uses more credits. In Gemini, higher effort uses more of your limit.
- Let it run, and answer its questions. Long Claude tasks keep running in the cloud if you close your laptop, and desktop tasks appear on your phone. Anything that needs local files or apps still needs Claude Desktop open.
- Review the output like a new colleague's work. Check sources, formulas and anything that will leave the building. See how to check if AI is telling the truth.
- Turn repeat jobs into a schedule. Claude's scheduled tasks can run in the cloud with your computer off, and Perplexity Automations run on a schedule or a trigger with review checkpoints.
A worked example: a weekly pipeline report that writes itself
A sales operations lead spends two hours every Monday turning CRM exports into a pipeline update.
- She puts the weekly exports in one folder and gives Claude access, with the permission setting on Manual for the first run.
- Her request: "Using the CSVs in this folder, build a spreadsheet of pipeline by stage and region with a chart, compare it with last week's file, and write a one-page update: what moved, what stalled, deals at risk and decisions needed. Flag any region down more than 10% week on week."
- Claude plans the steps, asks which column marks the close date, and returns the spreadsheet and the memo.
- She checks the totals against the CRM, fixes one stage mapping in her instructions, and switches the permission setting to Auto.
- She schedules the task for 8 a.m. every Monday. Her two hours become ten minutes of review.
Common mistakes with AI agents and how to fix them
- Vague goals. Agents work hard on the wrong thing. Fix: name the finished file and how you will judge it.
- Too much access too soon. Fix: start with one folder and Manual permissions, then widen.
- Letting it act on untrusted pages or emails. Text on a web page can try to give an agent instructions, which is called prompt injection. Fix: keep approval on for sending, paying and deleting, and be careful with agents that read your inbox.
- Running out of allowance mid-task. Perplexity's honest catch is that Computer runs on credits that can run out mid-task (active tasks pause and resume when credits are available), and Claude's heavier models use limits faster. Fix: start with a small version of the task.
- Skipping the review. Fix: treat every output as a draft until you have checked it.
Which AI agent to use
Pick Claude if your work lives in files and documents: it is the most complete agent inside an ordinary chat, with clear permission controls. Pick ChatGPT Work if you already use ChatGPT for everything and want agent work on the same plan. Pick Gemini if your day runs on Gmail, Docs and Drive, once Tasks reaches your account. Pick Perplexity Computer for research-heavy jobs where you want to choose models and effort. Anthropic also has Claude Duet, a proactive agent app in research preview for US Pro, Max and Ultra users through a waitlist, which asks before sending messages or paying.
Read the tool guides for Claude, ChatGPT, Gemini and Perplexity, and for app-to-app automation see best AI for automation and how to automate tasks with AI. Agents that click around your screen are explained in what is computer use, and the connector standard many use in what is MCP.
4 prompts for handing work to an AI agent
Prompts to copy
The outcome brief
Goal: [the finished thing, e.g. a spreadsheet and a one-page memo]. Material: [folder, files, websites or connected apps]. What good looks like: [criteria]. Rules: ask before sending, deleting or paying anything; mark anything you could not confirm. When you are done, list what you did, what you assumed and what I should check.
You get: A delegation brief any agent can follow
Plan first, then act
Before you start [task], write a short plan: the steps, the sources or files you will use, and any questions you need answered. Wait for my approval before carrying it out.
You get: A chance to fix the approach before the agent spends time or credits
Recurring report
Every [Monday at 8 a.m.], read the files added to [folder] since the last run and write a one-page update with these headings: [headings]. Keep the same structure each week and flag anything that changed by more than [threshold] since the previous report.
You get: A scheduled task that drafts a regular report
Review the agent's work
Review the output you just produced as a sceptical [manager]. Check every number against the source files, list any assumption you made, and tell me the three things most likely to be wrong. Fix what you can and mark the rest.
You get: A self-check before you do your own review