What is an AI browser? A browser agent can click, type and finish web tasks instead of only answering questions

Ask a normal AI assistant to find a restaurant and it may search the web and summarize options.

A browser agent can go further:

Open the reservation site
→ choose the date
→ set party size
→ compare times
→ open a restaurant
→ fill the form
→ ask for confirmation before the final action

The difference is action.

Search retrieves information; a browser operates an interface

Lesson 043 focused on finding and reading information.

A browser agent adds operations such as:

Read
+ Click
+ Type
+ Navigate
+ Submit

It can therefore act on a web interface rather than merely explain it.

What is a cloud browser?

A cloud browser runs in a remote computing environment rather than directly taking over every tab in your personal browser.

OpenAI’s current cloud-browser documentation, for example, describes a separate browser environment with its own cookies, login state and browser data rather than automatically inheriting your local history or saved passwords.

That separation is valuable: an agent can receive the minimum browser context needed for a task instead of your entire daily browsing environment.

Why not always use an API?

Lesson 018 explained that an application programming interface (API) gives software a structured way to communicate.

When a good API exists, it is often more reliable than imitating human clicks.

But many real workflows have only a website, hide functionality behind a front-end interface or require visual interpretation. A browser agent provides a more general fallback.

Connected apps are different again

A connected Gmail, Calendar or enterprise integration may expose structured tools directly.

Conceptually:

Connected app / API
→ structured and permissionable

Browser
→ very general, but dependent on changing user interfaces

Use structured integrations when they fit; use browser control when the interface itself is the only practical path.

The page itself can contain untrusted instructions

This reconnects to Lesson 035 on prompt injection.

A browser agent reads untrusted web content. A page could contain text designed to manipulate an AI system, such as asking it to ignore earlier rules or upload private data.

A secure design must distinguish:

• the user’s instruction,
• system policy,
• webpage content,
• tool permissions.

Consequential actions deserve confirmation

Searching and drafting are not equivalent to paying, deleting, publishing or submitting.

Good agent design separates low-risk preparation from irreversible or high-impact actions and asks for confirmation at the boundary.

A browser agent cannot bypass every website restriction

It can still encounter CAPTCHA, two-factor authentication, UI changes, popups, geographic restrictions or sites that block automation.

“Has a browser” does not mean “can complete every website task.”

One thing to remember

AI search mainly finds and reads web information. An AI browser can click, type and navigate. Once the model can act on a website, browser isolation, prompt injection defenses, permissions and confirmation become first-class design requirements.