#Agent web browsing

In the interview and in project chat, the agent has a browser of its own: a real Chrome that you can watch, take over and hand back. Use it to let the agent explore an existing site, or to show it how you do a task so it can write the requirements from what you did.

#Ask the agent to open a site

Ask in plain words, and say "in your browser": "Open https://app.example.com in your browser and look around the order screens." Without that, the agent may read the page as text, which shows nothing of an app built in JavaScript.

The first time the agent opens a page, its browser starts. That takes a few seconds, longer on a server that has not run one recently. An Agent browser panel then appears in the conversation. On a wide screen it sits to the right of the chat, half and half; on a narrow one it sits above the chat. The chevron beside it hides it.

Each conversation has its own browser. It closes shortly after the conversation's session ends.

#Take over

The panel starts in watch mode: you see what the agent does, and your mouse and keyboard stay yours. Press Take over to drive the browser yourself. The border turns orange and the panel reads You have control. Press Hand back when you are done.

Take over when the agent is stuck at a sign-in, a one-time code or a CAPTCHA, or when you want to show it how something is done.

#Show the agent how you work

While you have control, Flezise records what you do in the browser: the pages you open, what you click, what you type into each field and the forms you submit. It takes screenshots as you go. Password, token and one-time-code fields are recorded as ••••.

When you press Hand back, two things happen:

• The agent receives your steps as a numbered list in a message ("Opened /orders", "Filled “Customer” with “Acme”", "Clicked “Submit order”") and is asked to work out the flow, ask you about anything unclear and note the requirements it implies. In the interview and in project chat it can also open the key screenshots straight away. In the interview, the brief is written from that conversation.
• The steps are saved in the project's files as one reference document: a PDF with every step and up to twelve of the screenshots. The knowledge base indexes it within a few minutes, so later documents such as the PRD can draw on it.

The agent can also keep a page it finds worth remembering. Ask it to save the page, and it stores the page's screenshot, address and its note as one PDF in the same way.

#Keep a sign-in

After you sign in to a site while you have control, press Save sign-in. New browsers in the project then start signed in to that site. Forget sign-in removes it. The sign-in is kept as the site's cookies, so a site that keeps its session somewhere else asks you to sign in again.

#Browsing in coding runs

Coding runs, such as an implementation or its browser verification, use a browser inside their own sandbox instead. You do not see a panel for those.

#Next steps

• Agent sandboxes: where agents run and how their environment is set up.
• Knowledge base: how project files, walkthroughs included, are searched.