Anthropic Unveils Cowork: Revolutionary AI Agent Executes Computer Tasks!

Anthropic continues its pursuit of professional users. In a press release on their website, the company announced the launch of Cowork, a derivative of Claude Code designed to assist you with your work tasks. However, the company clarifies that this new tool is not just for developers. Here’s an introduction to it.

Cowork: An Assistant That Can Manage Your Files

When Cowork is activated, Claude can access the files you designate and operate directly on your computer within the boundaries you set. After recording your request, Cowork devises a plan and carries out the task progressively, keeping the user informed about ongoing actions. “If you have used Claude Code before, you will find familiar features: Cowork is built on the same principles,” Anthropic states. The tool can also utilize your existing connectors to access external information and can create documents and interact with the web when paired with Chrome.

For instance, Cowork can reorganize your downloads by sorting and renaming each file, create a new spreadsheet with a list of expenses from screenshots, or generate an initial version of a report from your scattered notes.

Cowork spares the user the need to manually provide context or adapt output formats, maintaining necessary elements throughout the tasks. It is also possible to initiate multiple tasks consecutively and let them run in parallel, while adding instructions or comments during processing.

Is Cowork Truly Secure?

The ability to deeply interact with the workstation naturally raises questions about data management, protection, and associated risks. Anthropic’s reassurances have been somewhat lacking in this regard.

While the company assures that Claude can neither read nor modify items that you haven’t granted access to and requires your consent before taking any significant action, it warns that “Claude can perform potentially destructive actions (such as deleting local files) if instructed to do so.” They add, “Since there is always a risk of Claude misinterpreting your instructions, you must provide very clear guidance on this.”

Regarding cyberattacks, Anthropic does not seem more confident. The startup discusses the risks of prompt injections, that is “attempts by attackers to alter Claude’s plans through content it might encounter on the Internet.” Despite efforts, the company admits that securing Claude’s actions “remains a significant development focus within the industry.” Therefore, it is advisable to avoid using the option for sensitive missions initially.

Limited Deployment for Now

Cowork marks a further step in the progression toward agent-based AI, a journey that Anthropic began with the launch of Claude 3.5 Sonnet over a year ago. In this field, the company will face competition from OpenAI and its Agent ChatGPT, now integrated into the Atlas browser.

However, for now, Cowork is only available as a research preview to subscribers of Claude Max through the macOS app.