Claude Cowork: The AI agent that built itself

Anthropic launched Claude Cowork on 12 January, and it’s basically Claude Code for non-developers. It can take on complex, multi-step tasks and execute them on your behalf - such as reading, editing and reorganising files in a folder. The interface sits in the Claude Desktop app and no terminal is needed.

You describe what you want done, and Claude plans the work, executes it step by step, and checks its own output. Anthropic describes the experience as 'less like a back-and-forth and more like leaving messages for a coworker'.

Claude Code has been one of Anthropic's breakout successes, but its developer-focused branding and terminal interface kept it out of reach for most users. Social media has been noticeably fizzing about Claude Code already, now Coworker has added more since last Monday.

Boris Cherny noted that people were already using Code for 'vacation research, building slide decks, cleaning up email, cancelling subscriptions, recovering wedding photos from a hard drive, monitoring plant growth, controlling your oven'. Cowork wraps those capabilities in something your non-technical colleagues might actually use.

As an example of its capabilities, Lenny Rachitsky fed Cowork 320 podcast transcripts and got back distilled themes in 15 minutes.

The most remarkable thing about it might be how it was made. According to Boris Cherny, Anthropic's head of Claude Code, the entire feature was built in roughly ten days, almost entirely by Claude Code itself. AI building AI tools. We're through the looking glass.

This is the shift from chatbots to agents that the industry has been promising for years. Programmer and blogger Simon Willison called it 'a general agent well positioned to bring the wildly powerful capabilities of Claude Code to a wider audience'.

Pietro Schirano, a former Anthropic engineer, predicted it would be 'a much larger disruption than anything else we've seen so far, because it's really going to impact the white-collar job'.

But not everyone's convinced. Claire Vo, founder of ChatPRD, said the outputs were just 'okay' and that Cowork 'exposes too much of its internal process for non-technical users'. Willison raised an eyebrow at Anthropic's security guidance, which advises users to 'monitor Claude for suspicious actions that may indicate prompt injection'. As he put it: 'I do not think it is fair to tell regular non-programmer users to watch out for that.'

Cowork is genuinely interesting, but the current caveats are significant. It's macOS only, with no web or mobile access. It's a 'research preview' limited to Max subscribers paying $100–200 per month.

Features like Projects and Memory don't work with it yet. And the security warnings are real: this thing can delete your files if it misinterprets your instructions, and prompt injection remains an unsolved problem.

Still, the story is pretty compelling. An AI tool that was largely built by AI, in under two weeks, that could plausibly automate chunks of white-collar work. That it’s essentially Claude Code with accessible UX is no bad thing. We can almost certainly expect something similar from OpenAI in the not-too-distant future.

Perhaps no surprises given a huge userbase, and huge losses: OpenAI announced it will begin testing ads in the United States for free and Go tier ChatGPT users. The ads will appear at the bottom of conversations and will be targeted based on discussion topics.

Users can dismiss ads, view explanations for targeting, and disable personalisation. Premium tiers including Pro, Plus, Business, and Enterprise will remain ad-free. OpenAI has committed to maintaining "answer independence," ensuring ads won't influence chatbot responses, and has promised not to sell user data to advertisers.

This was ‘The Thing That Matters in AI’ - a new edition of Absolutely Agentic that we will be sending every Monday at 1pm UK time during January.