as software WIRED reporter, I’ve tested a lot of shitty agents over the past few years. These experiences reveal a consistent pattern of generative AI startups over-promising and under-delivering when it comes to these “agent” helpers, programs designed to take control of computers and perform chores and digital errands, freeing up time for more important things. However, the bot you install on your laptop has a hard time completing even basic tasks. It just didn’t work.
This poor track record makes Anthropic’s newest agent, Claude Cowork, a pleasant surprise. When I tested it by running some basic and intermediate demos suggested by the company in addition to my own commands, it worked pretty well, especially for the software, which is still in beta. You can also organize your files into folders, convert file types, generate reports, take over your browser to search the web, and organize your Gmail inbox. When it comes to file management and computer interfaces, this tool feels like the beginning of a pleasant user experience evolution.
Over the past year, Anthropic has developed a cult following for its Claude Code tools among developers who love its ability to understand codebases and run commands. Technical staff all over San Francisco seem to be using this tool in their work all the time. But most of them aren’t members of the technical staff of hot startups.
“We played around with different ideas to see what form factors made sense for a less technical audience who didn’t want to use a device,” said Boris Cherny, Head of Claude Code at Anthropic. For the past two months, Cherny has been writing all the code with AI. Cowork was built using AI tools.
Cowork, which Anthropic released in research preview earlier this week, takes advantage of features available in the company’s coding-focused tools to make the user experience more approachable. This tool is designed for a wide group of non-technical users who want to try new ways to control their computers, but are uncomfortable with the command line.
Start
Reece Rogers
For now, Cowork is only available as part of a research preview to subscribers of Anthropic’s $100 per month plan. This is a common release strategy for generative AI companies that soft launch new features to early adopters.
Felix Rieseberg, a member of Anthropic’s technical staff who focuses on Cowork, says he uses Cowork to submit expense reports and convert files. “If this PDF is too large, make it smaller,” he says. “Convert these 20 JPEGs into one PDF. Report on all of them.” Although Rieseberg is excited that more advanced users are already experimenting with complex applications, he believes Research Preview’s “favorites” are intended for the simplest file-centric applications.
This initial release is limited to Claude on Mac and may be rolled out more broadly in the future. Also, although you can use it to work with files on your computer, Cowork requires an Internet connection to run.[Cowork]Tabs are available in the Claude app for macOS.[チャット]tab and[コード]Appears next to the tab. User sessions are labeled as “tasks” instead of “chat”.
What about security risks?
The biggest reason not to try Cowork is the ongoing security risks inherent in this type of agent. Like most agents, Cowork is susceptible to prompt injection attacks, secret messages hidden online that attempt to trick an AI tool into deviating from a task. Sensitive data should not be exposed to tools that can be compromised in this way.
Anthropic’s online support page says, “Claude can read, write, and permanently delete these files, so be careful when giving him access to sensitive information such as financial documents, credentials, and personal records.” They suggest storing backups of important files and creating a dedicated folder with non-sensitive information that you want Claude to have access to.
