Throughout this year, we are witnessing a fundamental shift in how we perceive the capabilities of large language models. While previously we looked forward to AI's ability to write a poem or explain a complex physical phenomenon, now the focus is shifting to agentic capability – that is, AI's ability to independently perform steps in a real environment. Anthropic, with the release of the Cowork tool, precisely confirms this trend.

From Programmers to Everyday Users: How the Inspiration for Cowork Arose

The path to creating Cowork did not begin in a laboratory inventing new algorithms, but by observing how people actually use existing tools. A year ago, Anthropic released Claude Code, a tool primarily intended for developers working in the terminal. Although it was a code-writing tool, Anthropic engineers noticed a fascinating phenomenon: people started using it for completely different purposes.

Boris Cherny, an engineer at Anthropic, stated on social network X that users were using this development tool to conduct vacation research, clean emails, and even attempt to control home appliances or organize photos from a hard drive. This "shadow work" using AI showed that users are not just looking for text answers, but a functional assistant that can manipulate data. The result is Cowork – a simplified version of Claude Code's capabilities, accessible even to people who have never heard of programming.

How the "Agentic Loop" Works and Interacts with Your Files

Unlike a typical interface where you have to copy text into a chat, Cowork operates on the principle of a sandbox. The user defines a specific folder on their computer to which Claude has access. Within this folder, the agent can perform complex operations.

The key to this technology is the so-called agentic loop. When you give the user a task – for example, "convert all receipt screenshots in this folder into a clear Excel table" – Claude doesn't just provide instructions on how to do it. Instead:

• Creates a plan: Breaks down the task into individual steps (browse files, recognize text, create a table).
• Performs the action: Actually opens the file, reads the data, and writes it into a new document.
• Checks the result: If it encounters an unreadable screenshot, it stops and asks you for clarification.

Examples of Practical Use

The possibilities for everyday work are wide. For a small business owner in the Czech Republic, it could be a tool for automatic invoice sorting. For a student, it could be about organizing lecture notes scattered across ten different text files. Anthropic itself gives examples such as reorganizing a chaotic "Downloads" folder or creating reports from fragmented notes.

Comparison: Claude vs. Competition

In the field of AI agents, Anthropic is entering a direct battle with giants like OpenAI and Microsoft. While Microsoft Copilot is deeply integrated into the Office 365 ecosystem, Cowork focuses on local work with your file system.

If we compare the capabilities of the Claude 3.5/4 (Opus) model with models like GPT-4o or Gemini 1.5 Pro, we see that Anthropic places greater emphasis on accuracy and "safe autonomy." While GPT-4o excels in speed and multimodal interaction, Cowork utilizes the specific architecture of the Claude Agent SDK, which is optimized for long, complex chains of tasks without losing context.

Price and Availability: Is it a Tool for Everyone?

Here comes the biggest obstacle for the Czech user. In its current phase, Cowork is part of a research preview and is exclusively available to subscribers of the Claude Max plan. The price of this plan ranges from 100 to 200 USD per month (which is approximately 2,300 to 4,600 CZK). This makes the tool primarily a product for power-users and corporate clients, not for the average end-user looking for a cheap alternative.

From an availability perspective, it's important to note that Cowork is currently only available for the macOS desktop application. Support for Windows or Linux has not yet been officially confirmed. However, the good news is that Claude as a model perfectly handles the Czech language, so you can enter instructions in Czech, and the agent will prepare the results for you in a Czech environment.

AI That Wrote Itself: A Fascinating Self-Reference

One of the most discussed aspects of Cowork's release is the way it was created. According to information from Anthropic's internal team, this entire tool was developed in an incredible ten days. And most importantly, the Claude Code tool was massively utilized for its development.

This process, where AI helps build other AI, creates a so-called recursive improvement loop. If this technology continues to accelerate, we can expect software development to speed up by an order of magnitude in the coming months, which will have a huge impact on the entire technological scene, including the European market, where the regulation of these autonomous systems within the AI Act is already being discussed.