agent-browser vs browser-use

TL;DR: agent-browser passed 15/15 runs, browser-use passed 12/15. agent-browser was also faster on every task both tools finished.

I wanted to know which browser tool to give my coding agent, agent-browser or browser-use. So I made both do the same five chores.

How I tested it

Codex drove both tools, so the model wasn’t a variable. Each tool got its own Brave window on the same Chromium build. Five tasks, three runs each.

I didn’t trust either tool’s “done”. After every run, a separate read-only connection checked what was actually on the page. That caught three bugs in my own checks first, which I fixed before scoring anything.

Commands vs code

This was the biggest difference. agent-browser takes plain CLI commands, one per action. browser-use’s CLI takes a Python script on stdin. Same form task, both ways:

agent-browser open https://www.selenium.dev/selenium/web/web-form.html
agent-browser type '[name=my-text]' 'Browser bench'
agent-browser type '[name=my-textarea]' $'Line one\nLine two'
agent-browser select '[name=my-select]' 2
agent-browser uncheck '#my-check-1'
agent-browser check '#my-check-2'
agent-browser click 'button[type=submit]'
agent-browser wait --text 'Received!'

browser-use <<'PY'
def click_name(role, name):
    nodes=cdp('Accessibility.getFullAXTree')['nodes']
    node=next(n for n in nodes if n.get('role',{}).get('value')==role and n.get('name',{}).get('value')==name)
    q=cdp('DOM.getBoxModel',backendNodeId=node['backendDOMNodeId'])['model']['content']
    click_at_xy(sum(q[0::2])/4,sum(q[1::2])/4)

goto_url('https://www.selenium.dev/selenium/web/web-form.html')
assert wait_for_load()
fill_input('[name=my-text]','Browser bench')
fill_input('[name=my-textarea]','Line one\nLine two')
click_name('combobox','Dropdown (select)')
press_key('Home')
press_key('ArrowDown')
press_key('ArrowDown')
press_key('Enter')
click_name('checkbox','Checked checkbox')
click_name('checkbox','Default checkbox')
click_name('button','Submit')
assert wait_for_element('#message',timeout=10)
print(page_info())
PY

For an agent, that means every browser-use step is a small program to get right. When a helper didn’t exist, my script had to build one out of raw CDP calls.

The tasks

Each clip is agent-browser doing the task.

01 · Read a table

agent-browser open https://the-internet.herokuapp.com/tables
agent-browser scrollintoview '#table1'
agent-browser get text '#table1 tbody'

02 · Fill a form

agent-browser open https://www.selenium.dev/selenium/web/web-form.html
agent-browser type '[name=my-text]' 'Browser bench'
agent-browser type '[name=my-textarea]' $'Line one\nLine two'
agent-browser select '[name=my-select]' 2
agent-browser uncheck '#my-check-1'
agent-browser check '#my-check-2'
agent-browser click 'button[type=submit]'
agent-browser wait --text 'Received!'

This is where browser-use failed, all three times. The page said “Received!” each time, but the data was wrong.

To be fair, part of this is on me. agent-browser has a select command. My browser-use script used arrow keys on the dropdown, which is what left it on the default. The missing line break is all browser-use though.

03 · Wait for content

agent-browser open https://the-internet.herokuapp.com/dynamic_loading/2
agent-browser click '#start button'
agent-browser wait '#finish'

04 · Switch tabs

agent-browser open https://the-internet.herokuapp.com/windows
agent-browser click 'a[href="/windows/new"]'
agent-browser tab t2
agent-browser wait --text 'New Window'
agent-browser tab t1

05 · Use a small app

agent-browser open https://demo.playwright.dev/todomvc/
agent-browser type .new-todo 'Buy milk'
agent-browser press Enter
agent-browser type .new-todo 'Read book'
agent-browser press Enter
agent-browser check '.todo-list li:first-child .toggle'
agent-browser click 'a[href="#/completed"]'

Both tools passed tasks 1, 3, 4 and 5 every time.

Results

Verdict

agent-browser. Plain commands instead of Python, faster per step, and no wrong form submissions. Stable tab ids and built-in recording are nice too. Every clip here came from its record command.

browser-use is fine at reading, waiting and tabs, and raw CDP access is handy if your agent already writes Python. I’d rerun the form with a proper native dropdown before calling 15/15 vs 12/15 final.