Anthropic just released Cowork, a powerful new tool that brings the agentic capabilities of Claude Code to non-developers, but the company is warning users to proceed with caution.
Cowork, which launched January 12 as a research preview for Claude Max subscribers, lets users give Claude direct access to folders on their computer. From there, Claude can read, edit, and create files autonomously to complete complex, multi-step tasks without having to code.
But here's the catch: Anthropic explicitly warns that Claude can take "destructive actions" such as deleting files if given unclear instructions. And early testers are already reporting a bumpy experience.
To look at what this means for the future of knowledge work, I turned to SmarterX and Marketing AI Institute founder and CEO Paul Roetzer on Episode 191 of The Artificial Intelligence Show.
The origin of Cowork traces back to a surprising discovery inside Anthropic's own offices. Boris Cherny, the creator of Claude Code, shared the backstory in a thread on X that's been viewed more than a million times.
Cherny recalls walking into the office one day and seeing a data scientist, not an engineer, using Claude Code to write SQL queries and generate plots for an analysis. Within a week, an entire group of data scientists had adopted it. Then designers started using it for prototypes. Finance used it for forecasting. Sales used it to analyze Salesforce data.
"We built Claude Code for engineers," Cherny wrote, "and here was a data scientist using it to do his work, too."
Eventually, people were using Claude Code to control ovens, recover wedding photos from broken hard drives, and analyze their own DNA. The pattern was undeniable: users wanted agentic AI for far more than coding.
"When we released Claude Code, we expected developers to use it for coding," Anthropic wrote in their launch announcement. "They did, and then quickly began using it for almost everything else. This prompted us to build Cowork."
And in a move that captures the moment we're in: Claude Code actually wrote all of the code for Cowork.
Cowork lives inside the Claude Desktop app for Mac. (And only on Mac at the moment.) You designate a folder, and Claude gains permission to work within it by reading files, creating new ones, editing existing documents, and executing multi-step workflows with minimal hand-holding.
Think of it as a non-technical, sandbox version of Claude Code. The interface removes the need for command-line tools or virtual environments. You just chat with Claude and let it handle the execution.
The tool also integrates with Claude's existing connectors, including Google Calendar or browser automation through Claude in Chrome, which opens up an even broader range of workflows. Anthropic has bundled in a set of "skills" that improve Claude's ability to generate specific file types such as Word documents and PowerPoint presentations.
Users are already testing it for tasks including organizing downloads, generating expense reports from receipt photos, drafting documents from scattered notes, and prepping for meetings by analyzing calendar data.
Not everyone's experience has been smooth. Claire Vo, founder of ChatPRD and host of the “How I AI” podcast, published a detailed breakdown after testing Cowork on several tasks.
Vo found that connectors didn't connect, terminal commands failed with "scary error messages," and local files wouldn't load properly. The tool asked for approval to open files too often, connected services she didn't request, and exposed too much of its internal workings (such as raw JavaScript files) to users who just wanted a final document.
"Working files are cool if you're a developer, but confusing if you're not technical," Vo wrote, adding, "Cowork right now sits in a fuzzy middle."
Still, she noted that despite its flaws, Cowork produced better outputs than a standard Claude chat, suggesting the underlying agentic planning does add real value.
What makes Cowork particularly notable is how openly Anthropic acknowledges the risks.
"By default, the main thing to know is that Claude can take potentially destructive actions such as deleting local files if it's instructed to," the company wrote in its announcement. "Since there's always some chance that Claude might misinterpret your instructions, you should give Claude very clear guidance."
They also flagged the risk of prompt injections, which are attempts by attackers to manipulate Claude's behavior through content it encounters online. Anthropic says it has built "sophisticated defenses," but admits that agent safety "is still an active area of development in the industry."
In other words: use at your own risk.
Beyond the immediate tool, Roetzer sees Cowork as an early signal of a much larger shift, one where AI agents work across entire business functions: marketing, sales, customer success,
Today, humans fill each of those roles. But Roetzer envisions a near future where a single human oversees a team of AI agents executing across all of them simultaneously.
“Develop the plan for all the go-to-market: The marketing plan, the sales plan, the customer success plan, all the documentation that needs to be created, all the email workflows that need to be built, the websites, the media plan, and then imagine just turning it loose," Roetzer says.
In that world, maybe the Chief Marketing Officer oversees the marketing agents, the sales leader oversees the sales agents, and so on. The humans don't disappear, but their role changes dramatically.
"All those things I just described is what the entry level people were doing," says Roetzer. "But if AI can do that now, all you really need is a director level person with experience and instinct and intuition."
This shift is already prompting questions about what new roles will emerge. Roetzer believes "orchestration" will become a critical skill, and possibly a formal job title.
"AI orchestrator, AI orchestration, AI ops orchestration…that idea of orchestration of all these agents might be something that you actually have a title for," he says.
He envisions a layered structure: project manager types who ensure agents follow processes and stay within guardrails, and director or VP-level strategists who evaluate whether the work makes sense, whether the strategy is sound, and whether the AI is generating ideas worth pursuing.
"You're going to need that level of project management ops," he says, "and then you're going to need someone asking:’OK, is the strategy sound? Is this the kind of thing I would build? Did it have new ideas?’"
For enterprise leaders, the immediate takeaway is clear: it's time to revisit your AI policies.
"If your employees are unclear if they're allowed to try this on their systems, you should be clear about that," says Roetzer.
Cowork might be early and imperfect, but it's a preview of what's coming from every major AI lab. OpenAI, Google, and others are all building toward a world where AI agents can take autonomous action on your behalf by reading your files, browsing the web, and executing complex workflows.
For knowledge workers, that means your job is about to involve a lot more orchestration and management of AI agents, not people.