Anthropic’s Claude Cowork: A Reliable AI Assistant

As a journalist at WIRED, I’ve encountered numerous disappointing AI products over the years—many promising more than they could achieve. These so-called agents were designed to manage our digital tasks to allow us more free time, yet they frequently failed to accomplish even the simplest functions.

In contrast, Anthropic’s new AI innovation is a refreshing deviation. During my evaluation, it efficiently handled both the standard and more complex scenarios suggested by the company and my own directions, performing admirably despite its beta status. Claude Cowork is capable of file organization, converting formats, generating summaries, and even managing web browsers and email applications. It represents significant progress in user-friendly file and system management.

A New Approach to Code and Beyond

Anthropic initially gained popularity with a code tool favored by developers, especially those in San Francisco. Recognizing that most people lack technical expertise, Anthropic explored various approaches to broaden its appeal.

According to Boris Cherny, Anthropic’s head of Claude Code, they experimented with different designs to suit users outside the technical community who are not fond of command lines. Cherny himself has adopted AI for all of his coding in recent months.

Introduction to Cowork

Cowork was recently unveiled in a research preview as part of Anthropic’s $100 monthly package, a common practice among AI start-ups introducing new features to early users.

Felix Rieseberg from Anthropic mentioned using Cowork for tasks like condensing PDF files and compiling JPEGs into single PDFs. These file-centric tasks have quickly become his preferred use of Cowork, although users are already experimenting with more complex applications.

Security Concerns

A major hesitance in adopting Cowork stems from inherent security vulnerabilities common to AI agents. Cowork, like any such tool, can fall prey to prompt injection attacks, where malicious hidden prompts attempt to mislead an AI's actions. Consequently, exposing sensitive information to such AI tools is risky.

Anthropic cautions users to limit Claude’s access by using dedicated folders with non-critical information and emphasizes the importance of file backups. They employ a virtual machine to restrict access strictly to specified folders.

Furthermore, if users permit, Cowork can interact with web browsers and email, but it comes with a notice alerting users to potential risks of code that might exploit system vulnerabilities.

Initial Impressions

I first tasked Cowork to manage the clutter of screenshots on my desktop. After allowing necessary permissions, it suggested sorting them by month and executed the task smoothly—an outcome I was pleased with, as it decluttered my digital space efficiently.

Subsequently, I experimented with the AI's ability to organize an email inbox. Although initial attempts were interrupted by technical snags, once instructions were specified to delete excess promotional emails, Cowork performed correctly, clearing out a thousand unwanted messages.

Lastly, I tested its ability to assist with more complex tasks. I connected Claude to my Google Calendar and requested assistance in booking movie tickets and scheduling the event. While it located the screening accurately, the system wisely refrained from completing the purchase autonomously. It successfully updated my calendar after manual confirmation.

Even in the face of bugs, Cowork emerged as the first AI agent to genuinely meet my practical needs. Developers aim to refine the tool further based on user experiences and feedback.