I Let an AI Agent Control My Browser for a Week. Here's What Nobody Tells You

I'll admit the first time I gave an AI agent control of my browser tab, I sat there with my hand hovering near the mouse the whole time, ready to grab it back. It felt less like using software and more like handing my car keys to a teenager and watching from the passenger seat. That feeling, it turns out, is pretty common right now, because this is genuinely new territory for most people, even ones who use ChatGPT or Claude every single day for writing and research.

What's different about these tools isn't that they can chat with you. It's that they can now click things. Open tabs, fill out forms, compare prices across ten sites, book something, summarize a page you haven't even scrolled to yet. That's the shift happening across the big three right now, and if you run any kind of business with a website, it's worth understanding what these agents can and can't actually do, because the marketing hype and the real experience are two very different things.

Who's doing what

OpenAI took its older Operator tool and folded it into something called ChatGPT Agent, and more recently launched ChatGPT Work, which is built for the kind of task that takes a human an hour or two rather than thirty seconds, think research projects with multiple steps, not just "find me a flight." Anthropic has been rolling out Claude for Chrome gradually since a small pilot last year, and it's now available across their paid plans. Google went the widest with Gemini in Chrome, making it generally available for Workspace users and adding what they call Auto Browse for Pro and Ultra subscribers, which started reaching phones only a few months ago.

So all three exist now, in some form, inside a browser. But they're not really built for the same person.

The part everyone skips over

Here's the thing I wish someone had told me before I spent a weekend testing all three: these agents are not close to as reliable as the demos make them look. There's a benchmark called OSWorld 2.0 that tests these agents on realistic, multi-step computer tasks, the kind of thing you'd actually ask them to do. The best performer in the entire test, using the most capable setup available, still only fully completed about one in five tasks. One in five. Everything else was a partial success, a wrong turn, or the agent quietly doing something adjacent to what you actually asked for.

That's not a knock on the technology. It's genuinely impressive that it works at all a fifth of the time on tasks that involve real judgment calls across multiple websites. But it does mean the honest way to use these tools right now is as an assistant you double check, not one you hand the keys to and walk away from. I learned this the hard way when I asked an agent to compare pricing across a few SaaS tools and it confidently gave me a summary that mixed up which plan included which feature. It sounded completely right. It wasn't.

Picking the one that actually fits you

If you're a normal person who wants help summarizing a long article, comparing a few products before buying, or pulling information together from a handful of open tabs, Gemini in Chrome or ChatGPT Agent are the more natural starting points. They're built to feel like a personal assistant sitting next to your browsing, not a separate tool you have to learn.

If you're more technical, if you're debugging your own website or trying to build some kind of automated workflow into a product you're making, Claude for Chrome tends to be the one people reach for, especially alongside Claude Code. It's less "chat with your browser" and more "give precise instructions and watch it execute them," which suits people who already think in steps and want control over exactly what happens.

Neither answer is really about which company is "winning." It's closer to picking a tool for a job. Asking which one is best is a bit like asking whether a screwdriver or a wrench is better. Depends entirely what you're trying to tighten.

What this means if you run a website

There's a quieter implication in all this that I think gets missed. If these agents are increasingly the thing browsing on someone's behalf, comparing your pricing page against a competitor's, summarizing your blog post instead of the person reading the whole thing, then how your site is structured starts to matter in a slightly different way. A page that's clear, that states things plainly instead of burying the actual answer under three paragraphs of fluff, is easier for both a human and an agent to get right. A confusing page doesn't just lose a human reader anymore. It can also get summarized badly by whatever's browsing on that person's behalf, and you don't get a say in how.

I don't think that means rewriting your entire site around robots. It just means the old advice, write clearly, say the useful thing first, stop hiding the answer behind clickbait, matters for a slightly different reason than it used to.

Where I've landed after using all three

Honestly, none of them replaced anything I do manually yet. What they've replaced is the small, annoying research tasks I used to do half-heartedly anyway, the fifteen browser tabs I'd open and never fully compare. For that, they're genuinely useful. For anything that actually matters, a booking, a purchase, something with real consequences if it gets it wrong, I'm still watching the screen with my hand near the mouse. Given that even the best one gets it fully right one time in five, that seems like the sane way to use it for now.