Anthropic on Monday released Cowork, a new AI agent feature that extends the power of its hugely successful Claude Code tool to non-technical users. According to insiders, the team built the entire feature in about a week and a half, primarily using Claude Code itself.
The announcement marks a major shift in the race to bring working AI agents to mainstream users, and positions Anthropic to compete not only with OpenAI and Google in conversational AI, but also with Microsoft’s Copilot in the burgeoning market for AI-powered productivity tools.
"Cowork lets you complete non-technical tasks just like developers use Claude Code." The company made the announcement through X’s official Claude account. This feature is available as a research preview exclusively to Claude Max subscribers (Anthropic’s $100-$200/month power user tier) through the macOS desktop application.
Over the past year, industry talk has focused on large-scale language models that allow you to write poetry and debug code. With Cowork, Anthropic is betting that the real company value lies in AI that can open folders, read messy piles of receipts, and generate structured expense reports without a human being involved.
How a developer who uses coding tools for vacation research inspired Anthropic’s latest product
Cowork’s origins lie in Anthropic’s recent success with the developer community. In late 2024, the company released Claude Code, a terminal-based tool that allows software engineers to automate rote programming tasks. The tool was a hit, but Anthropic noticed a strange trend where users were forcing coding tools to do more than just code.
According to Boris Cherny, an engineer at Anthropic, the company has observed that users are deploying developer tools for an unexpectedly wide variety of tasks.
"Since we launched Claude Code, we’ve seen people using it for all sorts of non-coding tasks. Research your vacation, create slide decks, clean up your email, cancel subscriptions, recover wedding photos from your hard drive, monitor plant growth, control your oven," Charney wrote to X: "These use cases are diverse and surprising. The reason is that the underlying Claude Agent is the best agent and Opus 4.5 is the best model."
Recognizing this use of shadows, Anthropic effectively removed command line complexity from developer tools and created a consumer-friendly interface. In a blog post announcing the feature, Anthropic explained the developer: "I immediately started using it for almost everything else," which one "This is an easier way for anyone to work with Claude, not just developers."
Inside a folder-based architecture that allows Claude to read, edit, and create files on your computer
Unlike a standard chat interface where users paste text for analysis, Cowork requires a different level of trust and access. The user specifies specific folders on the local machine that Claude can access. Within that sandbox, the AI agent can read and modify existing files or create entirely new files.
Anthropic provides several examples. Reorganize a messy downloads folder by sorting and intelligently renaming each file, generate an expense spreadsheet from a collection of receipt screenshots, create a report from notes spread across multiple documents, and more.
"Cowork allows Claude access to folders on your computer. Claude can then read, edit, or create files in that folder." The company explained about X. "Try creating a spreadsheet from a pile of screenshots or a first draft from scattered notes."
This architecture is called "agent loop." When a user assigns a task, the AI does more than just generate a text response. Instead, develop a plan, perform steps in parallel, check your work, and ask for clarification when you hit a roadblock. Users can queue multiple tasks and have Claude process them simultaneously. This is a workflow that Anthropic describes as “feeling.” "Even more so, it’s not like reciprocating, it’s like leaving a message for a colleague."
The system is built on Anthropic’s Claude Agent SDK and shares the same underlying architecture as Claude Code. Anthropic points out that Cowork: "It can take on many of the same tasks that Claude Code can handle, but in a more approachable format for non-coding tasks."
The recursive loop in which AI builds AI: Claude Code reportedly wrote much of Claude Cowork
Perhaps the most notable detail about Cowork’s launch is the reported speed at which the tools are being built, highlighting the recursive feedback loops in which AI tools are being used to build better AI tools.
In a livestream hosted by Dan Shipper, Anthropic employee Felix Rieseberg confirmed that his team built Cowork in about a week and a half.
Alex Volkov, who is in charge of AI development, expressed surprise at this timeline. "Holy crap, Anthropic just built “Cowork” in… a week and a half?!"
This immediately led to speculation about how much of Cowork itself was built by Claude Code. Simon Smith, vice president of generative AI at Klick Health, puts it bluntly about X: "Claude Code wrote everything about Claude Cowork. Can we all agree that we are at least somewhat in a recursive improvement loop here?"
The implications are serious. Anthropic’s AI coding agent could have made a significant contribution to building its own non-technical sibling. If true, this is one of the most visible examples to date of AI systems being used to accelerate their own development and expansion. This strategy could widen the gap between AI labs that are successful in bringing their own agents in-house and those that are not.
Connectors, browser automation, and skills extend the reach of Cowork beyond your local file system
Cowork doesn’t work alone. This feature integrates with Anthropic’s existing ecosystem of connectors, tools that link Claude to external information sources and services like Asana, Notion, PayPal, and other supported partners. Users who configure these connections in the standard Claude interface will have them available within their Cowork sessions.
Additionally, Cowork can be combined with Anthropic’s browser extension, Claude for Chrome, to perform tasks that require web access. This combination allows agents to navigate websites, click buttons, fill out forms, and extract information from the Internet while operating from a desktop application.
"Cowork includes a number of novel UX and safety features that we think make this product truly special." Charney explained emphatically. "Built-in VM (virtual machine) for isolation, out-of-the-box browser automation support, support for all claude.ai data connectors, and ask for clarification when in doubt."
Anthropic also introduced the following initial sets: "skill" Designed specifically for Cowork, it powers Claude’s ability to create documents, presentations, and other files. These are built on the Skills for Claude framework, which the company announced in October, and provides a specialized set of instructions that Claude can load into specific types of tasks.
Why Anthropic warns users that its AI agent may delete files
The transition from chatbots suggesting edits to agents making edits comes with significant risks. An AI that can organize files could theoretically delete them.
Anthropic demonstrated remarkable transparency, devoting significant space in its announcement to warning users about the potential dangers of Cowork. This is an unusual approach for a product launch.
The company specifically acknowledges that Claude does the following: "May take potentially destructive actions (such as deleting local files) if directed to do so." Because Claude can sometimes misinterpret instructions, Anthropic urges users to provide the following information: "very clear guidance" Regarding sensitive operations.
Of further concern is the risk of prompt injection attacks. A prompt injection attack is a technique in which a malicious attacker could embed hidden instructions in content that Claude might encounter online, allowing the agent to bypass security measures or perform harmful actions.
"We have built advanced defenses against instant attacks." Antropic writes, "However, agent safety, the task of protecting Claude’s real-world behavior, remains an area of active development in the industry."
The company characterized these risks as being unique to the current state of AI agent technology, rather than specific to Cowork. "These risks are not new to Cowork, but they may be the first to use more sophisticated tools that go beyond simple conversations." Presentation notes.
Anthropic’s desktop agent strategy directly challenges Microsoft Copilot
The launch of Cowork puts Anthropic in direct competition with Microsoft. Microsoft has been trying to integrate its Copilot AI into the fabric of its Windows operating system for years, with mixed results.
However, Anthropic’s approach differs in its insularity. By restricting agents to specific folders and requiring explicit connectors, we seek to balance the utility of OS-level agents with the security of sandboxed applications.
Anthropic’s approach is characterized by bottom-up evolution. Rather than designing an AI assistant and refurbishing agent capabilities, Anthropic first built a powerful coding agent, Claude Code, and is now abstracting that functionality for a broader audience. This technology lineage could allow Cowork to achieve more robust agent behavior from the start.
Claude Code has generated a lot of excitement among developers since it was first launched as a command-line tool in late 2024. The company expanded access through its web interface in October 2025, followed by Slack integration in December. Cowork is the next logical step. The idea is to provide the same agent architecture to users who never touch a device.
Who can access Cowork now and what’s next for Windows and other platforms?
For now, Cowork is limited to Claude Max subscribers using the macOS desktop application. Users in other subscription tiers (Free, Pro, Team, Enterprise) can join the waitlist for future access.
Anthropic has demonstrated clear intent to expand the scope of this feature. In this blog post, the company explicitly mentions what it learned from the research preview and its plans to add cross-device sync and bring Cowork to Windows.
Cherny set expectations appropriately, describing the product as follows: "It’s similar to how Claude Code felt when it was first released."
To access Cowork, Max subscribers must download or update the Claude macOS app and "Cowork" It’s in the sidebar.
The real issues facing enterprise AI adoption
For technology decision makers, the impact of Cowork extends beyond a single product launch. The bottlenecks for AI adoption are changing. Model intelligence is no longer the limiting factor, but rather workflow integration and user trust.
As the company says, Anthropic’s goal is to make working with Claude feel more like delegating to a colleague than operating a tool. Whether mainstream users are ready to hand over folder access to an AI that can misinterpret instructions remains an open question.
However, Cowork’s speed of development (with major capabilities built in 10 days, perhaps by the company’s proprietary AI) portends a future in which the capabilities of these systems will compound faster than organizations can appreciate.
The chatbot learned how to use the file manager. You never know what you’ll learn to use next.
