Which computer use agent is best currently?
After testing four of the many available, one stands out as the best, for now
Ever since Anthropic announced a computer use agent last fall, many companies and open source groups have been working furiously on similar projects.
A "computer use agent" can operate software on a computer to accomplish a task you give it, allowing you to work on other things while it churns away. Some of them can only operate a web browser, while others can operate any software installed on your computer.
I've tested four of the many systems currently available:
My real-life test cases
I've given all of these tools a couple of real-life test cases. First is to find a low-cost round-trip airline ticket from San Francisco to Orlando in the near future so I can visit my parents. It's the kind of routine research one may perform on the web when thinking about taking an optional trip when you might have time off from work but want to base it upon low airfares and an approximate and flexible date range.
Another test is to help manage my fantasy baseball roster for the day. I've been in a league with one of my sons and several of his friends for 9 years. Since there are games every day of the season, it takes time to keep your roster updated with good players who are starting that day.
Which one is best?
For the most part, these tools are fascinating to watch, but slow and error-prone. Like captions in a movie, they announce each step and retry in deep detail, like "I see a calendar, let me choose the date the user asked me about" or "I see a sorting icon at the top of the list, I'll click that to sort by date". But they often misunderstand visual elements or get stuck when something pops up in the user interface, obscuring a control they need to move forward.
And by slow, I mean at least 10 times slower than a human would perform a task.
But of these, one stands out as reasonably fast and accurate: Perplexity Comet.
It's actually just a browser (it can't use other locally installed apps on your computer) but does remarkably well. If you've used Perplexity, it is easy to use Perplexity Comet and really fun to watch it conduct its work to accomplish the task you give it. And because it is reasonably fast, the process of iterating its work with refinement instructions isn't frustrating like it is with the other tools.
A glimpse of things to come
I think this is just a glimpse of the future, where we'll increasingly use web browsers less and less, and instead ask AI agents to accomplish tasks for us, and it will use software to do our bidding. Who would have thought, just a few years ago, that we'd witness a reduction of web browser usage starting around now?
