Agents, and when to supervise them

An agent does not just answer, it acts. The more it can change, the closer you watch.

From answering to doing

Most of this course is about Claude answering questions. An agent is different: it takes actions for you. It can click through websites, fill in forms, open files, move things between apps and finish multi-step jobs while you do something else.

Claude offers agents in two main shapes:

Claude in Chrome: a browser extension that can browse, read pages, fill forms and complete tasks on websites you are signed in to.
Cowork: Claude works through longer tasks with your files and apps, reporting back as it goes. You can start, steer and review a task on the web, in the desktop app or on your phone.

Both are powerful. Both make mistakes. The question for every task is not “can the agent do this?” but “what happens if it does this wrong?”

The three-zone rule

Sort every agent task by what it can change.

Zone	What the agent can do	What you do	Examples
Green: read only	Look, gather, summarise	Let it run, check the result	Collect prices from five supplier sites; read a folder of PDFs and list the deadlines
Amber: reversible changes	Draft, create, reorganise things you can undo	Watch, approve each step or review before anything is saved	Draft replies in your inbox; rename and sort files into folders; fill a form but stop before submit
Red: irreversible or external	Send, pay, delete, post, sign	Do it yourself, or have the agent stop and hand over at the final step	Send an email to a customer; place an order; delete records; post to social media; agree to terms

An agent should never be in the red zone on its own. Not because it will always fail, but because when it fails you cannot take it back.

The three zones at a glance:

Three zones for agent tasks: green read-only runs alone, amber reversible changes are watched, red irreversible actions stay with a person — Sort by what the agent can change, not by how clever the task is.

What goes wrong

Agents fail in ways that look like success:

Wrong target. It fills the form perfectly, on the wrong customer’s record.
Confident guess. A field it cannot find gets a plausible value instead of a question.
Scope creep. Asked to tidy one folder, it tidies three.
Instructions on the page. A website can contain hidden text telling an AI to do something. Good agents resist this; none are perfect. This is a reason to keep agents away from payments and credentials.

How to brief an agent

Say exactly what done looks like. “Draft replies to the three unread emails from suppliers, do not send.”
Put a boundary in words. “Only touch the folder called 2026 Invoices.”
Ask it to stop before the irreversible step. “Fill the order form and stop. I will click submit.”
Watch the first run of any new task. Let it run unsupervised only after you have seen it succeed.

A brief that has all four parts in it:

A sample agent brief with four numbered parts: what done looks like, do not send, a folder boundary, and stop before submit; with a reminder to watch the first run — Every sentence in the brief does a job: defines done, forbids sending, limits scope, hands over before submit.

Anthropic’s demonstration of Claude for Chrome working through tasks in a browser. Notice which steps you would put in each zone:

Video: Let Claude handle work in your browser · Anthropic · 2 minWatch for the moments where the agent reads, where it drafts, and where a person would need to approve before anything is sent.Watch on YouTube (opens in a new tab)

How it looks today · CoworkChecked Nov 7, 2026

Cowork is available on claude.ai (web), in the Claude desktop app and in the Claude phone apps, on paid plans (in beta on Team; on Enterprise an owner turns it on for web and phone). On the web, start from the Home tab and pick Cowork in the message box; if your account has the newer combined Claude experience there is no separate Cowork option, you just describe the task in any conversation. On Pro, Max and Ultra new tasks run in the cloud, so they keep going when your laptop is closed and you can pick them up on another device.

Access to the folders on your computer, local connectors and browser use still goes through the desktop app, and only while it is open. You point Cowork at a folder or connected app, describe the job, and it works through the steps, showing a running log and pausing when it needs a decision. Check which folders and connections you have granted before starting; it can only touch what you have allowed, so allow the minimum.

Menus and button names change often. If your screen looks different, the skill is the same: look for the equivalent option.

And the equivalent for Cowork on the desktop:

Video: Getting started with Claude Cowork · Claude · 4 minWatch for how folders and apps are granted up front, and how the running log shows each step as it happens.Watch on YouTube (opens in a new tab)

Try it now

Write down the next three tasks you would like an agent to do. Put each in a zone: green, amber or red. For any red task, rewrite it so the agent stops at the final step and hands over to you.

Remember: read-only tasks can run alone, reversible tasks need a watcher, irreversible tasks stay with a person. Watch the first run of anything new.

Checking your progress…