WorkBuddy's Browser Use Lets AI Drive Your Logged-In Browser Without a Script
Browser use has been largely blocked for developers behind China's network and overseas-payment hurdles; WorkBuddy puts a scriptless, logged-in browser agent on the domestic desktop for free. The catch applies everywhere: an agent that improvises its path and silently swaps its success criteria is fine for exploratory work but cannot carry regression suites that demand repeatability and explicit validation.
Tencent's WorkBuddy, the AI office workbench that crossed 20 million monthly visits in July, now ships browser use through a Chrome, Edge, or QQ browser extension. Instead of writing Selenium or Playwright locators, you describe a task in one sentence and the agent opens pages, clicks buttons, scrolls, and collects results inside your existing logged-in browser. In tests, it autonomously navigated an interview question bank, scrolled through list pages, and executed a three-step sequence on an already-open tab after a confirmation prompt.
The feature deliberately stays out of your way: it works in a separate tab group with its own cursor, and a mobile app lets you assign tasks to your computer from anywhere. But it is not a replacement for scripted automation. Each run takes minutes, the path can differ between runs, and it has no assertions, so critical-path verification still belongs to tools like Playwright. The division of labor that emerges: hand it one-off, low-frequency tasks such as post-release smoke tests, online inspections, and competitor research; keep stable, high-frequency regression on scripts.
The silent fallback is a double-edged sword: it keeps tasks moving when data is unavailable, but it changes the evaluation standard without disclosure, so anyone relying on the output must audit the basis, not just the result.
The central claim—that the 'valuable thing is the sentence you use to assign the task'—is more than rhetoric: agent reliability depends on precise goals and acceptance criteria, because a vague prompt gets replayed as a vague execution.
The feature's nondeterministic path each run is the real boundary: it separates exploratory and ad-hoc work from anything requiring reproducibility, so browser use will not displace scripted automation for stable, repeatable processes anytime soon.