|
Written by: Ronnel DG |
The AI Agent can read your website's code, but reading code isn't the same as using a website. The agent-browser skill lets the agent open your site in a real browser, click through pages, menus, and forms, and report the problems a visitor would run into. It works on your published site, on your staging site (the private preview address where you check changes before publishing), and on other public websites.
Note: You need to add the skill before the agent can browse anything. Each browsing request runs like any other AI Agent request, so it uses AI credits.
What the browser skill does
On its own, the agent works from your site's source code. It can see that a contact form exists and that its fields are wired up, but it can't confirm that pressing Submit sends anything. It can see that a page exists, but not that the page loads correctly for a visitor.
With the browser skill added, the agent does what you would do. It loads the page, clicks the buttons, fills in the fields, and watches what comes back. Then it reports the result in the chat, along with pictures of what it saw.
This is different from pointing the agent at a web address. Referencing an external website reads the content of a page so the agent can borrow ideas or copy details from it. The browser skill interacts with the page instead: it moves through a site and finds out whether things work.
Adding the browser skill
The browser skill comes from the same catalog as B12's other skills, so adding it takes three clicks.
- Click
Menuat the top of the B12 Editor. - Select
Agent
Skills - Click
Add skill, then click agent-browser in the list.
The skill is added straight away and stays with this website. If you manage more than one B12 website, add it to each one separately. For everything else skills can do, see Using and creating AI Agent skills.
Checking your own website
Once the skill is added, you don't need to name it. Ask for the outcome you want and the agent picks the skill up on its own.
Testing a form or sign-in flow
Anything a visitor has to fill in is worth testing, because a form that looks right in the editor can still fail when someone uses it. Tell the agent which flow to try and what a good result looks like:
Open my website and try to [submit the contact form on my Contact page]. Tell me whether it worked, and show me anything a visitor would see go wrong.
The agent walks through the flow, then reports what happened at each step.
Tip: Say what success looks like, not only what to click. "Check that the thank-you message appears after I submit" gives the agent something to confirm. "Check my form" leaves it guessing.
Choosing your staging or published site
The agent can browse either version of your site. Your staging site is the private preview where your latest changes sit before you publish. Your published site is what the public sees, which may be a few changes behind.
Name the one you want so there's no doubt:
Check the [staging] version of my site and click through every page in the main navigation. List any page that fails to load or shows placeholder text.
Testing on staging first is usually the safer order. You find the problem, fix it, and only then publish.
Reviewing other websites
The agent can also browse public websites that aren't yours. A common use is keeping an eye on a competitor, so you can see how your own pages compare.
Visit [competitor web address] and take screenshots of the products they sell. Summarize what they offer and how it differs from my site.
The agent can reach any page a visitor could reach without signing in. Pages behind someone else's login, and anything a site blocks to automated visitors, are out of reach.
A check like this is more useful repeated than done once. You can put the same request on a schedule and have the agent run it monthly, which is covered in Scheduling tasks for the AI Agent.
What the agent sends back
Browsing results arrive in the chat in two forms. Screenshots show a single moment, which is what you want for a layout problem or an error message. Screen recordings show the whole run, so you can watch the agent move through a flow and see exactly where it went wrong.
Recordings are short, usually a few seconds per flow, and they capture the agent's own session rather than any visitor's.
Nothing on your site changes while the agent is browsing. It's looking, not editing. When you want a problem fixed, ask for it in your next message, and that request saves a version of your site as usual.
Frequently asked questions
→Does browsing my website use AI credits?
⇒Yes, based on how much work the run involves. If you put a browsing check on a schedule, the credits come out when it runs, not when you set it up.
→How is this different from asking the agent to reference a website?
⇒Referencing a website reads what's on a page so the agent can use the content. The browser skill uses the page: it clicks, fills in forms, and reports what happened.
→Can the agent fix the problems it finds?
⇒Yes, but not in the same breath. Browsing only looks at your site. Ask for the fix in your next message, and the agent makes the change.
→Can it get to pages that need a sign-in?
⇒The agent will require credentials in order to access any gated pages or content on your website. While we don't recommend sharing credentials directly in the chat, for testing purposes, you can create a test account and provide these credentials to sign in with. Don't provide a real user's information. We recommend removing the test credentials after testing or rotating the password regularly for your security. On someone else's website, the agent can only reach pages that are open to the public.
→Does it work before I publish?
⇒Yes. Ask it to check your staging site and it tests your latest changes, before anyone else can see them.