Anthropic Just Launched Cowork — And It Might Be the AI Agent That Actually Works
Anthropic Just Launched Cowork — And It Might Be the AI Agent That Actually Works
Last week, while scrolling through my usual tech news feeds, I stumbled across something that made me pause. Anthropic — the company behind Claude — quietly dropped a new tool called Cowork. No flashy keynote. No Super Bowl ad. Just a research preview tucked into the Claude Desktop app for Mac. And after spending a few days testing it, I’m genuinely impressed.
Here’s the thing about AI agents: most of them don’t work. We’ve been promised a future where autonomous bots handle our emails, organize our files, and book our meetings. Instead, we got chatbots that hallucinate meeting times and “productivity assistants” that require more setup than actual work. Cowork is different.
What Exactly Is Claude Cowork?
Cowork is essentially Claude Code — Anthropic’s wildly successful coding tool — repackaged for people who don’t write code. If you haven’t been living under a rock, Claude Code has been the breakout AI product of the past year. It reached $1 billion in annualized recurring revenue less than a year after its debut, and by the end of 2025, that number had grown by another $100 million. Claude Code now accounts for roughly 12 percent of Anthropic’s total ARR, which sits around $9 billion.
But here’s the catch: Claude Code requires you to be comfortable with a terminal. You type commands, the AI executes them, and you review the results. That’s great for developers. It’s a non-starter for everyone else.
Cowork removes that barrier entirely. Instead of a command-line interface, you get a conversational tab right next to Chat and Code in the Claude Desktop app. You tell it what you want done in plain English, and it handles the rest — organizing files, managing emails, browsing the web, even booking event tickets.
“We tried a bunch of different ideas to see what form factor would make sense for a less technical audience that doesn’t want to use a terminal,” Boris Cherny, Anthropic’s head of Claude Code, told WIRED. The result is a tool that takes everything Claude Code can do and makes it accessible to literally anyone.
What Can Cowork Actually Do?
I tested Cowork with several real-world tasks, and here’s what I found.
File organization. My desktop was a disaster — screenshots, PDFs, random downloads spanning months. I gave Cowork access to my desktop folder and asked it to organize everything. Before taking any action, it asked my preferences for how to sort the files, recommending a folder-per-month structure. About a minute later, my desktop was clean. Every file sorted into the right place. No manual drag-and-drop. No third-party cleanup app.
Email management. I connected Cowork to my Gmail and asked it to clean up my inbox. The tool asked what my goals were and what types of clutter to focus on. It auto-generated some suggestions — archive old newsletters, delete promotional emails older than 30 days, flag messages from specific senders. My first attempt to batch-archive promotional emails hit some snags, but when I pivoted to asking it to simply delete a thousand unread promotional messages, it clicked through and did the job in about two minutes.
Real-world task completion. This is where Cowork got interesting. I connected it to Google Calendar and asked it to find two tickets to an evening showing of a movie at a theater near me, then add that event to my calendar as a date night. It found a 9 PM showing, pulled the details, and added the calendar entry — complete with a reminder. Something a human assistant would handle in five minutes, done by an AI agent without me lifting a finger.
Felix Rieseberg, a member of technical staff at Anthropic who focuses on Cowork, shared his own use cases: filing expense reports, converting file formats, shrinking oversized PDFs, and combining multiple JPEGs into single documents. “Turn these 20 JPEGs into one PDF. Make me a report about all of these things,” he said. The kinds of tedious tasks that eat up hours every week for knowledge workers.
The Security Question — Because It Matters
Let’s be honest: giving an AI agent permission to read, write, and permanently delete files on your computer is terrifying. And Anthropic knows it.
“Since Claude can read, write, and permanently delete these files, be cautious about granting access to sensitive information like financial documents, credentials, or personal records,” Anthropic’s support page warns. The company recommends saving backups of critical files and creating a dedicated folder for Cowork to work in.
The architecture underneath is designed to limit damage. Cowork uses a virtual machine under the hood, which means you have to explicitly choose which folders Claude can access. If you don’t grant permission to a folder, Claude literally cannot see it. This isn’t a theoretical privacy guarantee — it’s a technical boundary enforced at the system level.
Cowork can also interact with your browser, but with important caveats. It asks for permission at each step — reading a website, clicking a button, filling a form. The browser interface includes an explicit disclaimer about hidden content and potential prompt injection attacks, where malicious websites could try to trick the AI into doing something unintended.
Cherny says Anthropic designed Cowork with multiple safety mitigations, including prompt-injection detection, keeping users in the loop on every action, and using virtualization to restrict file access to specifically requested folders only. These aren’t perfect protections, but they’re more thoughtful than most AI tools ship with.
Why This Matters for the AI Industry
Cowork represents something significant: the moment AI agents start moving from developer tools to mainstream products. Claude Code proved that developers would pay serious money for AI that can actually do work — not just suggest autocomplete, but understand entire codebases, write features from scratch, and debug complex issues. Cowork asks whether the same model works for non-technical knowledge workers.
The market timing is notable. Claude Opus 4.5 — Anthropic’s latest model — appears to have been the inflection point. “The only model I can point to where I saw a step-function improvement in coding abilities recently has been Claude Opus 4.5,” says Kian Katanforoosh, an adjunct lecturer on AI at Stanford and CEO of Workera. “It doesn’t even feel like it’s coding like a human. You sort of feel like it has figured out a better way.”
That same capability — understanding context, executing multi-step tasks, handling ambiguity — is what powers Cowork. The underlying technology that made Claude Code a $1 billion product is now being applied to the billions of office workers who’ve never opened a terminal in their lives.
We built the simplest possible thing. The craziest thing was learning three months ago that half of the sales team at Anthropic uses Claude Code every week. — Boris Cherny
Cherny’s point about Anthropic’s own sales team using Claude Code is telling. If non-technical employees at the company building the tool are adopting it organically, there’s a real product-market fit signal there. Cowork extends that adoption curve to an even wider audience.
The Limitations You Should Know About
Cowork is not perfect, and Anthropic is upfront about that. Currently, it’s only available on Mac as part of a research preview for subscribers of Anthropic’s $100-per-month plan. That’s a steep price for a tool that’s still in beta and has known limitations.
In my testing, Cowork occasionally stumbled on batch operations — particularly with email management, where it sometimes got confused about which messages to process. It’s also worth noting that Cowork requires an active internet connection to run, despite operating on local files. The AI processing happens in the cloud, not on your machine.
The prompt injection risk is real and shouldn’t be minimized. Any AI agent that can browse the web and interact with applications is vulnerable to hidden instructions embedded in websites or documents. Anthropic’s mitigations are solid for a first release, but this is a fundamentally hard problem that the entire industry is still working on.
And then there’s the broader philosophical question: as these tools get better, what happens to the work they’re automating? Cherny addressed this directly: “As an industry, we’ve always gone through transitions. My grandfather was a programmer in the Soviet Union, and he was programming on punch cards. It was a very physical thing. And then at some point, it turned into machine code, and then it turned into the first high-level languages, like C.”
Who Should Try Cowork Right Now?
If you’re already paying for Anthropic’s $100/month plan, Cowork is worth experimenting with today. The file organization alone — taking a chaotic desktop and structuring it in under a minute — is something I’d pay for separately. The email management is less polished but already useful for bulk cleanup tasks.
If you’re not a Claude subscriber, the math is harder. $100/month is a lot for a research preview. But if you’re a knowledge worker spending hours every week on file management, inbox cleanup, data entry, or report generation, Cowork might already be worth the price of admission — assuming you’re on Mac.
What I’d really like to see is a Windows version, a standalone pricing tier that doesn’t require the full $100/month subscription, and continued improvements to the email and browser automation features. These are coming — Anthropic says it plans to keep updating Cowork based on user feedback — but for now, the tool is what it is: a surprisingly capable first release from a company that has a track record of shipping well-engineered products.
The Bottom Line
Cowork is the first AI agent I’ve used that actually does what it promises. Not perfectly, not without occasional hiccups, but well enough that I’d trust it with real tasks. That’s a higher bar than you might think — I’ve tested dozens of AI “productivity” tools over the past three years, and most of them were glorified chat interfaces with a task list bolted on.
The significance of Cowork isn’t just in what it does today. It’s in what it signals: the best AI companies are no longer building chatbots. They’re building agents that can take action in the real world — in your files, in your inbox, on your calendar. And the ones that do it well, with proper security guardrails and a focus on user control, are going to define the next era of personal computing.
Claude Code proved developers were ready for AI agents. Cowork is Anthropic’s bet that everyone else is too. After using it, I’m inclined to think they might be right.
📖 Related: Google Meet’s AI Note-Taker Is Breaking Out of the Virtual World
📖 Related: Anthropic Launches Cowork: The Claude Desktop Agent That Works in Your Files Without Any Coding
📖 Related: OpenAI Now Lets Teams Make Custom Bots That Can Do Work on Their Own
