Is Anthropic Cowork Ready for Everyday Work?

TL;DR

Anthropic Cowork can automate useful multi-step work, but it currently requires substantial setup, patient debugging, and close supervision. Testing showed that it could eventually build a daily vocabulary email system and reduce time spent summarizing AI funding deals, yet errors, missing features, terminal troubleshooting, and a steep learning curve make it unsuitable as a seamless tool for most general users.

Transcript

In the last [music] few weeks, Anthropics Co-work has captured considerable attention. [music] And no wonder Co-work has a crazy promise that it can apply [music] all the power of Anthropics coding agent, Claude Code, [music] to non-programming tasks in a simple tool that anyone can use. But here's what a deep [music] dive investigation revealed th... Read More

Key Insights

• Cowork is a chatbot-style adaptation of Claude Code intended to bring coding-agent capabilities to non-programmers. It can write code, edit files, and interact with web pages, but its more limited computer access is designed to reduce concerns about accidental file deletion.
• Cowork was built by Anthropic engineers in about two weeks after users began applying Claude Code beyond programming. The source describes people using Claude Code to prepare tax filings, book theater tickets, and plan vacations, which helped demonstrate demand for a more approachable interface.
• The vocabulary automation test exposed Cowork's current usability problems. A detailed 352-word prompt led to stalled responses, an unwanted return from Opus to Sonnet, confusing setup questions, and roughly an hour of debugging through the terminal before the first email arrived.
• The first vocabulary email was incomplete despite two hours of effort. Some definitions were missing or incorrect, and the requested example sentences were absent, showing that completing a workflow does not guarantee the output satisfies all stated requirements.
• Early adopters reported that Cowork can provide real value on unpleasant or repetitive tasks. One tester used it to file thousands of dollars in expense reports and reorganize an email inbox, while another successfully generated 200 possible titles for a research paper.
• Cowork can overlook important information when handling broad prioritization tasks. When asked to prioritize a research team's assignments, it initially recommended dropping important work and later produced a supposedly comprehensive list with gaps because it had focused only on the top 20 tasks.
• The recommended Cowork learning path starts with a simple action, then progresses to workflows involving multiple tools. Anthropic's Boris Churnney suggested beginning with an email response before attempting document analysis, spreadsheet creation, discrepancy detection, and sending results through Slack.
• Cowork currently works best as a supervised collaborator rather than a plug-and-play autonomous worker. The reported results improved through repeated guidance, eventually producing a functioning vocabulary system and helping cut the expected workload for AI funding summaries by about half.

Questions & Answers

Q: What is Anthropic Cowork designed to do?

Anthropic Cowork is designed to apply capabilities associated with Claude Code to non-programming and multi-step work through a familiar chatbot-like interface. It can write code, edit files, and interact with web pages without theoretically requiring users to work primarily through a terminal. Its computer access is also more limited than Claude Code's, which may reduce concerns about accidental file deletion.

Q: How did Cowork perform in the vocabulary email test?

Cowork eventually produced a working system that emailed vocabulary words every morning, but reaching that result required extensive effort. The first attempt took about two hours and included stalled responses, model switching, confusing server questions, and terminal debugging. Its initial email also had missing or incorrect definitions and omitted the requested example sentences, so additional back-and-forth was necessary.

Q: Why did the reporter need to use the terminal with Cowork?

The reporter needed the terminal while setting up the email server recommended by Cowork. About an hour was spent pasting Cowork's suggestions into the terminal and returning error messages to Cowork for further debugging. This contradicted an important part of Cowork's appeal, because its chatbot-like interface was supposed to make advanced agent work more accessible to people uncomfortable with command-line tools.

Q: What useful tasks did early Cowork users complete?

Early users completed several practical tasks with Cowork. One doctoral candidate used it to file thousands of dollars in expense reports and reorganize an email inbox. A research team leader asked it to generate 200 possible titles for a paper, which it handled well. The reporter also used it to automate vocabulary emails and summarize daily AI funding deals.

Q: What problems appeared when Cowork prioritized team assignments?

Cowork initially advised a research team leader to drop some important assignments. After it later presented what it called a comprehensive list, the user discovered gaps. When questioned, Cowork acknowledged that it had considered only the top 20 tasks. The episode shows that users must verify claims of completeness and check whether the agent has silently narrowed the scope of its work.

Q: How should new users learn to work with Cowork?

Anthropic's Boris Churnney recommends starting with a simple task, such as responding to an email. Users can then progress to work requiring multiple tools, such as reviewing financial documents, creating a spreadsheet of discrepancies, and sending a Slack message about the results. The final stage is running several Cowork projects simultaneously, a capability that many users may not realize is available.

Q: Who is Cowork currently best suited for?

Cowork is currently best suited for early adopters who have a specific use case, realistic expectations, patience, and some technical comfort. The reporter suggested it may work reasonably well for someone with coding experience whose skills have become rusty. General users expecting an effortless tool may struggle with debugging, unclear questions, incomplete outputs, and the time required to learn the system's quirks.

Q: Should most people use Cowork now or wait?

Most people may be better served by waiting a few more months unless they have a specific task and enough technical comfort to manage setup and mistakes. Cowork can already create value, as shown by the working vocabulary system and faster AI funding summaries, but it remains a research preview with a steep learning curve, inconsistent results, and a need for close supervision.

Summary & Key Takeaways

• Anthropic built Cowork from Claude Code in about two weeks to make agent capabilities accessible through a chatbot-like interface. Cowork can write code, edit files, and interact with web pages while having more limited computer access than Claude Code. It is offered as a research preview through Claude's desktop app for paying Mac users.

• A reporter tested Cowork by asking it to rebuild an unreliable vocabulary email system containing more than 800 collected words. The initial attempt involved stalled responses, model switching, confusing questions, and about an hour of terminal debugging. After two hours, the resulting email still contained missing definitions, incorrect information, and no requested example sentences.

• Later experiments produced more encouraging results. Cowork eventually created a working system that emailed vocabulary words every morning and helped summarize daily AI funding deals, potentially cutting a four-hour weekly task in half. The broader conclusion was that Cowork resembles a capable intern who needs onboarding, correction, technical comfort, and carefully scoped assignments.