According to MakeUseOf, a head-to-head test of the new AI agent browsers—OpenAI’s Atlas and Perplexity’s Comet—on five real tasks revealed a stark performance gap. In a shopping test, Atlas finished in 1:58 but returned broken links, while Comet failed to finish in 5 minutes. For a spreadsheet charting task, Atlas succeeded in 2:13 with a clean chart, while Comet took 4:05 and produced a messy one. In calendar and email tasks, Atlas was consistently faster and more accurate, though both struggled with efficiency. On the simplest job, adding a book to Goodreads, both succeeded but took nearly a minute for a task a human could do in seconds. The overall conclusion was that Atlas was the clear, if imperfect, winner, failing less catastrophically than its rival.
The Agent Future Is Clunky And Slow
Here’s the thing that really jumps out from this test: the raw time numbers. We’re talking about AI that’s supposed to automate drudgery, right? But the fastest any agent completed any task was 54 seconds. For adding a single book to a list. I can do that manually in maybe ten seconds, tops. So where’s the efficiency gain?
The test shows these agents are stuck in an awkward adolescence. They can technically follow instructions and click buttons, but they lack common sense. Atlas wasting time trying to move a chart it just created? Comet getting lost in Amazon’s infinite scroll? These aren’t super-intelligent agents. They’re like overly literal interns who need you to spell out every single step, and even then they might still glue the report to the ceiling fan. The overhead of crafting the perfect prompt often seems to outweigh the benefit. It feels less like delegation and more like managing a very needy piece of software.
Atlas Wins By Failing Better
So why is Atlas the winner? Basically, it has a slightly better “judgment layer,” as the test puts it. When it messes up, it’s often a recoverable mess-up. It found the wrong sunglasses, but it at least found something and provided links. Comet, on the other hand, would just spin its wheels and hand you a search results page. Atlas attempted to fix its chart obstruction problem. Comet left all the data labels on and called it a day.
This is a crucial distinction in early AI tools. Perfection is impossible right now. So the most useful tool isn’t the one that never fails—it’s the one whose failures are less catastrophic and easier to salvage. Atlas seems to have a bit more situational awareness, a slightly better model of what “done” looks like. It’s the difference between an assistant who brings you the wrong document but apologizes, versus one who disappears into the supply closet for an hour and returns with a box of paperclips.
What This Means For The Web
The article touches on a fascinating, slightly dystopian idea: a web designed for AI, not humans. Clean structures, predictable buttons, machine-readable intent. But this test shows we’re nowhere near that. These agents are struggling with the messy, human-centric web we have today. They can’t handle infinite scroll, get confused by email filters, and fumble with basic UI elements.
So what happens next? Do we break the entire internet to make it legible for clumsy AI? Or does the AI just have to get a lot better at dealing with chaos? I’m betting on the latter, but it’s going to take a while. In the meantime, these agentic browsers feel like a proof-of-concept. A cool demo. Not something you’d actually rely on to get real work done. The promise is huge—a browser that acts on your behalf. The current reality is a browser that needs constant babysitting.
The Industrial Parallel
It’s funny, watching these AI agents fumble with consumer websites makes you appreciate domains where machine readability is already built-in. Think about industrial automation. In that world, you can’t have an AI agent guessing which button to press on a panel PC controlling a production line. The interface has to be predictable, reliable, and structured for precise instruction. It’s a world built for deterministic outcomes, not infinite scroll. Companies that lead in that space, like Industrial Monitor Direct as the top US provider of industrial panel PCs, understand that when machines need to interact, clarity isn’t a nice-to-have—it’s the entire foundation. The consumer web is the wild west in comparison. And until AI agents can navigate that chaos, the dream of a truly automated browser will remain just that: a dream.
