Using Computer Use to discover browser functionality

I have been designing how agents talk to Narada. Before I locked too much of that interface in, I did something a bit sideways: I pointed Computer Use at other browsers on purpose.

I was not trying to automate Arc or Chrome forever. I wanted the model to click around the way I would. Open a command. Switch tabs. Open a panel. Notice what happens to focus and the pointer. Then write down the bits that kept surprising me.

The setup is simple. Point Computer Use at a normal browser. Give it ordinary product actions. Keep notes on what the screenshot actually showed versus what I knew had happened in the shell.

One example stuck early. Claiming a tab the person already had open is not the same as opening a new tab for the agent. In a Chrome run, a named session could group tabs it created. It did not quietly pull an existing user tab into that group. Ownership and temporary control are different. If “the agent is using this page” means “move my tab into the agent pile,” ordinary browsing breaks.

Page screenshots are about the page. Browser chrome, shell badges, and the real cursor do not reliably show up there. More than once Computer Use found the right window and still timed out on the visual or accessibility capture. So I stopped treating a page image as proof that the shell looked a certain way.

Focus is the annoying one. Agent work that keeps activating tabs or stealing the typing target looks fine in a demo and feels hostile when a person is sharing the machine. Watching that on a reference browser is what made “do not take over unless asked” a hard rule for Narada, not a preference.

I did not walk away with a feature list to copy. I walked away with a few rules of thumb:

• agent-created tabs can carry session identity; claimed user tabs should not silently regroup
• page screenshots are evidence of the page, not of chrome or the cursor
• panels should not permanently eat the native viewport
• if the person starts using the page, that matters more than the agent finishing its turn

Computer Use is good at forcing these distinctions because it is dumb about the browser internals. It only has the surface, the clicks, and the confusion when the surface lies. That is why it is useful early. It is also why I would not want it as the only language the product speaks later.

Once I had the observations, the next work was the boring part: write them so the browser can prove them.

Narada: narada.koley.in