Anthropic has equipped AI with a computer to make it do real work.

The Claude Cowork product launched by Anthropic and its core feature, Dispatch, demonstrate how AI can transform from a conversational tool into a genuine productivity collaborator by having an independent local virtual machine environment. The development team achieved this innovation in just ten days by allowing the AI to autonomously write most of the code, proving that in the context of significantly reduced execution costs, "practice first, then screen" has become a new R & D paradigm. This system allows users to create and combine automated workflows through natural language instructions and can handle complex long - process tasks such as file management, code repair, and schedule scheduling. The author also points out that this technological leap is profoundly affecting the job market, especially as entry - level positions face the risk of being replaced, forcing career development and education models to accelerate transformation. In short, equipping AI with a dedicated computer not only improves work efficiency but also marks that **Artificial General Intelligence (AGI)** is reshaping the underlying logic of social operation.

On March 17th, Claude Cowork launched a new feature: Dispatch.

You assign a task on your phone and then leave. When you come back, the task is done: the files are organized, the tables are generated, and the process is completed.

Throughout the process, no one is operating the computer.

Behind this is a brand - new attempt by Anthropic: directly providing Claude with a dedicated computer.

Felix Rieseberg, who is in charge of the Cowork product, talked about three things in a recent interview: why this design was chosen, how it was rushed out in 10 days, and what impact it would bring.

Section 1 | AI Needs Its Own Computer

Anthropic provided Claude with a computer.

It's not for it to operate your computer but to configure an independent working environment for it in a virtual machine.

Felix Rieseberg said:

"Programmers need a computer to write code. If the company doesn't provide a computer and they can only send code via email, they simply can't work."

The same goes for Claude. If it can only answer you in the dialog box but can't access files, install tools, or connect to the browser, no matter how smart it is, it can't do much.

So Anthropic provided Claude with a Linux virtual machine. It can install Python, Node.js, call the browser, and read and write files.

This virtual machine is not large but has complete functions. Claude can install the necessary tools, run code, and process data without asking you for permission every time. More importantly, it is an isolated environment. Your local files will not be accessed randomly, and Claude's operations are within controllable limits. It has enough freedom to complete tasks but also has clear boundaries.

Why choose local instead of the cloud?

Felix's reason is:

"The entire Silicon Valley has underestimated the value of local computers. Why do we all use laptops instead of just tablets or cloud devices?"

Because many work scenarios cannot be solved by the cloud.

If you synchronize all your work data to the cloud, first there is the issue of permissions: you have to authorize each cloud tool separately, and managing them is a disaster.

Secondly, there is the security issue: many websites will directly lock your account when they detect that you log in from two different places. Banks are especially strict.

More importantly, local operation means that the data does not leave your computer. For enterprise users, this is a necessity.

Anthropic's choice is to provide Claude with its own virtual machine, which runs on your local computer. The data does not leave your device, but Claude has enough ability to complete complex tasks.

The newly launched Dispatch feature is based on this design: you assign a task to Claude on your phone and then leave. It will complete the task on your computer, and you can just come back to see the result. Claude will call local files, connectors, and various tools and finally hand in the tables, documents, and processing results.

This allows many processes that originally needed to be completed manually to be handed over in whole.

Felix himself is using it: he asks Cowork to go through the system crash logs, automatically filter out the bugs that can be fixed, and then schedule multiple Claude instances to make modifications separately. The entire process does not require manual intervention.

Some users also use it to process video materials: they throw a batch of files to Cowork, and it automatically organizes, renames, and generates descriptive copy, completing the entire publishing process. There is no need to click manually one by one.

These things cannot be done by the AI in the dialog box. Because it can't access your file system and can't run a long - process continuously.

But once it is given a computer, everything is different.

It changes from "a tool that can answer questions" to "a collaborative partner that can take over tasks."

Equipping AI with a computer is to enable it to really do work.

Section 2 | Claude Wrote 90% of Its Own Code

Equipping AI with a computer sounds like a huge project.

But Anthropic completed this in just 10 days.

Most of the code for Cowork was written by Claude itself.

Felix said in the interview that Anthropic's engineers rarely write code manually. Their work is more about telling Claude what to do, and then Claude is responsible for implementation.

In fact, Claude now writes about 90% of Anthropic's codebase. In some development sprints, Claude almost completed the entire codebase.

Felix said:

"We don't write memos now. We just make all the candidate solutions and then pick the best one."

In the past, when developing a product, one had to first sort out the requirements, write specifications, evaluate technical solutions, and then develop. Each step required a lot of time and manpower. Now, after the execution cost has become lower, many decisions can skip the discussion and start directly. Can't decide whether Plan A or Plan B is better? Then make both A and B and decide based on the results.

Because the cost of "making it" itself has become very low.

When the development cost is low enough, product development changes from "think before doing" to "do first, then choose."

Cowork itself was developed in this way.

It has a core feature: Skills.

What are Skills? They are instruction files written in plain text, telling Claude how to complete a certain type of task. The design is very simple: write a description in Markdown format, telling Claude which tools to call, what process to follow, and what details to pay attention to for this task.

That's it. There is no complex configuration, no code, and you don't even need to understand technology.

Felix said: "Anyone can create a Skill. A text message might be a Skill."

More importantly, Claude can generate Skills by itself.

A user's approach is to first let Cowork process a batch of files and complete the entire uploading process. After that, tell Cowork: "Turn the previous process into a Skill."

Cowork did as told.

Then he found that this Skill was too large, some steps might fail, or sometimes only a certain part of it was needed.

So he told Cowork again: "Split this Skill into three smaller ones."

After splitting, Cowork also automatically generated a parent Skill to coordinate these three child Skills.

Throughout the process, the user didn't write a single line of code, just used natural language to say what to do.

Some people also use it to manage their calendars: every morning, Cowork automatically checks the schedule, reschedules when there are conflicts, and reminds when there are important meetings. This is also a Skill, and it can be continuously adjusted according to personal habits.

These scenarios have a common point: let the AI take over an entire process. And the more you use it, the better it gets, because more and more Skills suitable for your work scenarios will be accumulated, which can be reused, shared, and combined.

It's like playing a factory - building game: start by automating some small things, then keep adding, and finally the entire work process will be optimized.

Cowork was developed in 10 days using this method.

The Anthropic team directly made several candidate solutions, tested, compared, and selected. Claude took on most of the code - writing work, and humans were responsible for setting the direction, making decisions, and controlling the quality.

What Cowork proves is: With AI, the way of doing things has changed. It has changed from thinking clearly before doing to doing first and then choosing.

Section 3 | There Isn't Much Time Left for Adaptation

When AI starts to really do work, the first ones to be affected are those who are just starting their careers.

Felix said bluntly:

"At Anthropic, we are very worried about the impact these tools will have on the job market in the future, especially on entry - level employees."

Tasks such as organizing documents, processing data, running processes, and doing preliminary analysis, which are repetitive and standardized, are originally assigned to new people in many industries.

Because this is how new people accumulate experience.

Now, these tasks are starting to be replaced by AI.

On the other hand, the efficiency of experienced people is increasing. After using such tools, senior engineers' output has increased significantly. They can hand over basic work to AI and focus on the parts that require more judgment themselves.

The gap between the two ends is widening:

Experienced people double their efficiency because they know how to break down tasks and how to schedule tools.

Newcomers find that the basic work they used to learn from is disappearing.

The market impact will be quite large, and people are not ready yet. Career paths may need to be redesigned.

How to redesign? The University of Waterloo provides an idea. The graduates of this university are more prepared in many companies. The reason is that the courses include a large number of internships, and students intern in different companies to accumulate real - world work experience.

Felix observed: "Graduates with internship experience are obviously more psychologically prepared than those who only study in the classroom."

When basic work is compressed, the entry path also needs to change. The Waterloo model provides a direction: let new people contact real - world work scenarios earlier.

In the interview, the host also proposed another idea: Can we use AI to simulate these scenarios?

The logic is: In traditional work, it may take you three years to experience a few truly growth - promoting moments. For example, leading a project for the first time, handling a major bug for the first time, and reporting to a customer for the first time. There may only be a few dozen such moments that allow for rapid growth, and most of the other time is spent on repetitive tasks.

What if we use AI to create a simulated environment for these key scenarios, allowing people to experience decision - making, making mistakes, correcting, and summarizing in virtual projects? In this way, in a week, you can basically gain the equivalent of three years of project and experience.

Just like going to college, people pay to learn how to do things. It's just that this time it's more urgent, and the learning process is greatly compressed.

Because there isn't much time left for adaptation.

Many people are asking: When will AI surpass human intelligence? That is, when the capabilities of AI accelerate to a critical point and start to self - reinforce, and the way the entire society operates will be changed.

Some say it may take another 10 years, while others say it may only take 1 year.

Felix's judgment is:

"It doesn't make much sense to argue whether it's five or ten years now. If we are quite sure that AI will be smarter than humans, we should quickly take action."

That is to say, the direction of AGI has been determined.

The emergence of products like Cowork is a sign of this direction.

The impact it brings will affect career paths, education methods, and the structure of the entire labor market.

These problems cannot be solved by a single company and require the entire society to respond.

Conclusion

AI won't replace humans all at once.

It first takes over repetitive work.

When execution becomes cheaper, deciding what to do is more crucial than just doing it blindly.

Reference Materials:

https://www.youtube.com/watch?v=ZpZ7lFoWaT8&t=1s

https://www.latent.space/p/felix-anthropic

https://www.linkedin.com/posts/felixrieseberg_quick-personal-update-after-two-really-rewarding-activity-7325655275984637956-v0zu

Source: Official Media/Online News

This article is from the WeChat official account "AI Deep Researcher", author: AI Deep Researcher, editor: Shen Si. Republished by 36Kr with authorization.