Website logo
Home

Blog

Claude Cowork, we tested the first artificial intelligence agent that works properly | Italy

Claude Cowork, we tested the first artificial intelligence agent that works properly | Italy

The new Anthropic tool finally manages files, browsers and calendars in a satisfactory way.Here is our first impression of Claude Cowork As a journalist, I have tested many bad AI agents over the past few years.Artificial intelligence startups, on the...

Claude Cowork we tested the first artificial intelligence agent that works properly  Italy

The new Anthropic tool finally manages files, browsers and calendars in a satisfactory way.Here is our first impression of Claude Cowork

As a journalist, I have tested many bad AI agents over the past few years.Artificial intelligence startups, on the other hand, tend to make more promises than they deliver when it comes to these "staff" assistants, programs designed to interact directly with the computer and perform digital tasks and tasks for us, giving time to more important things.In practice, however, the bots we use Installed on my laptop are difficult to complete basic tasks.It didn't work.

Questo bilancio poco incoraggiante rende l’ultimo agente di Anthropic, Claude Cowork, una piacevole sorpresa. Quando l'ho messo alla prova, il sistema se l'è cavata piuttosto bene, soprattutto per un software ancora in versione beta. Cowork è in grado di organizzare file in cartelle, convertire formati, generare report e persino prendere il controllo del browser per fare ricerche sul web o riordinare una casella Gmail. Sul fronte della gestione dei file e delle interfacce del computer, questo strumento sembra l’inizio di un'evoluzione positiva dell’esperienza utente.

Last year, Anthropic already cultivated a kind of "cult" following the faith of Claude Code, a system highly valued by developers for its ability to understand all codebases and execute commands, which quickly established itself among professionals in San Francisco.However, it must be said that many people do not have what it takes to become the first employee on the crest of the wave in Silicon Valley.

"We tried a few different ideas to see what kind of factor makes sense for a less technical audience, who don't want to use a terminal," explains Boris Cherny, head of cloud code at Anthropic.Over the past two months, Cherny has increasingly relied on AI for programming, and Cowork itself has been built with the help of artificial intelligence tools.

Released by Anthropic on January 12 as a research preview, Cowork takes the capabilities of the company's tool and makes it accessible.It is designed for a general audience who may be interested in trying a new way to manage computers but are not comfortable with the command line.

First steps with Cloud Colleague

Cowork is currently only available in search preview as a subscription to Anthropic's $100 monthly plan. This is a fairly common release strategy among AI companies, which introduce new features in a controlled manner starting with an initial group of subscribers.

Felix Reseberg, one of Anthropic's co-workers at Cowork, says he uses an AI agent to manage payment information and transaction information.Rieseberg is surprised how advanced users are already experimenting with more complex applications, but he likes the simple, focused way of using the data.

Although wider distribution is planned for the future, CoWork is currently only available in the Mac de Clod version.Although the tool can interact with files on your computer, it requires an Internet connection to use.In the Clod app for MacOS, the CoWork tab appears next to Chats and queues, and user sessions are referred to as "Tasks" rather than "Chats".

Security risks

The main reason you might not try Cowork is the security risks inherent in AI agents.Like many similar tools, Cowork is vulnerable to prompt injection attacks, hidden online messages designed to trick AI systems and sabotage their behavior.It is best not to provide sensitive data to tools that can be compromised in this way.

"Because Claude can read, write, and permanently delete these files, use caution when granting access to sensitive information such as financial documents, credentials, or personal records," states Anthropic's online support page.The company recommends saving backup copies of important files and creating a dedicated folder containing only non-sensitive information that Claude has access to.

"We're using a virtual machine," Cherney explains."This means you have to specify which folders Claude can access. If you don't give him access to a folder, Claude won't see it."

If authorized, Anthropic's new AI agent can also navigate your browser to search for information or help manage email.The bot asks for permission for many of its actions, such as reading web pages, making it clear that hidden code on the Internet portal can "steal data, install malware on systems, or control them."

Cowork, designed by Anthropic to incorporate several security measures, including instant injection detection, uses mechanisms designed to inform users of what agents are doing and the tool's use of "virtualization," which allows access to only specific files, Cherny said.

The first attempt I have been to, the information if Claode can set up a photographic article settled in a desk, including embarrassment.To do this, I need to give it to the entire entry and allow it to correct the files (as it will end up with them if necessary).

Before the chatbot asked my preferences, a database shows a single digit of the procedure in a desk, marked by the field.

I tested a series of systems by running larger and more complex queries, such as setting up an inbox.After granting access, Cowork asked me what my goal was for Gmail and what it should focus on.For each question, the tool provides additional buttons with automatically generated answers.

When I asked the agent to save the emails instead of deleting them, the first attempt had some problems because the bot had difficulty managing the storage of the large number of some promotional messages.Against this error, I changed my strategy and asked Cowork to immediately delete a thousand unread emails.At this point, the tool clicked here and there and deleted what I asked to delete, without touching anything else.Of course, there's always the real possibility that Cowork could make a mistake, as Anthropic continues to work on bugs and perfect the new feature.

Finally, I connected the cloud to Google Calendar and asked Cowork to find two tickets to an evening screening of Marty Supreme at a theater near me and then add the event to the calendar as a date night.A slightly more complicated operation with economic consequences.The bot saw a show at 9pm but stopped before buying tickets for security reasons.After I completed the purchase manually, the calendar updated.

Bottom line, Cowork is not perfect in its current state and its developers will continue to update it based on user feedback.That said, I can say that this is the first AI agent that has really convinced me.

Explore daily updates and news including top stories in Sports, Tech, Health, Games, and Entertainment.

© 2025 SDI Online, Inc. All Rights Reserved.