Theo reviews Anthropic’s new Claude Co-Work, arguing it’s basically Claude Code repackaged for non-developers: a useful, sandboxed computer agent for file work, browser control, and other real-world tasks. He likes the product concept and thinks it hints at a broader shift toward AI doing actions, not just generating text, but he spends much of the video criticizing its rough UX, confusing permissions, and weak app quality.
Watch on YouTube ›Get the market thesis, key claims, assets, contradictions, and follow-up questions from any financial video — then unlock a version personalized to your portfolio, watchlist, and favorite speakers.
Theo’s core thesis is that Claude Co-Work is a genuinely interesting product because it moves Claude from “answering questions” into doing practical work on a computer, but that the current implementation is clumsy enough that Anthropic is not yet proving it can ship great end-user software. He repeatedly says the product makes sense, calls it “our first taste of AGI in a very real way,” and frames it as a normal-person UI wrapped around capabilities he already uses in Claude Code. At the same time, he is blunt that the desktop app and login flow are buggy, confusing, and often frustrating in ways that undercut the product’s promise. A big part of his praise is based on his own day-to-day use of Claude Code for non-code tasks: analyzing iMessage history via SQLite, organizing files, and locating content in a downloads folder. …
Near term, this looks like a product narrative trade: the launch can attract attention fast, but usability bugs and confusing permissions could cap adoption until the onboarding is smoothed out.
Over the next few months, the setup improves only if Anthropic turns Co-Work into a dependable workflow for file-heavy and browser-heavy tasks; otherwise it remains a demo-led feature rather than a habit-forming product.
Longer term, the important shift is toward agentic operating-system layers where AI can act on files, browsers, and local apps; the winners will likely be the companies that pair strong models with trusted execution and excellent UX.
Claude's agent safety features cannot guarantee safety against prompt injections and jailbreaks.
The speaker critiques Anthropic's wording around agent safety, saying they should have been more direct about the inability to guarantee safety.
Prompt injections in AI agents will not be taken seriously until there is a high-profile incident.
Speaker observes that the industry needs a publicly visible security event before prompt injection risks get proper attention.
Claude's co-work feature was inspired by the popularity of the open-source Claudebot project over the holiday break.
Speaker notes that the Claude Code team built co-work in 2 weeks using Claude Code, implying inspiration from Claudebot's popularity.
What does co-work look like and how does it work?
The co-work tab sits next to existing chat and code tabs in the Claude desktop app. You start with a prompt, optionally attach a folder, and it begins working. Simon tested it on his blog drafts with a prompt about checking which drafts weren't published and suggesting ones closest to being ready.
What is Anthropic Labs and why was co-work created?
Labs is Anthropic's attempt to incubate products and projects internally to build user-facing experiences. Co-work happened because someone internally wanted Claude Code-like capabilities, built it, and before the company could shut it down it got too popular. Labs is an internal program to make more things like this happen — an incubator approach common at big companies but new to Anthropic.
Why can't co-work figure out how many iMessage messages you got in the last 30 days?
Co-work initially couldn't do it, but when switching to Claude Code via terminal it could query the iMessage database and found 4546 messages. The difference is that co-work runs in an isolated VM environment and requires the user to manually select the folder containing the data — it can't just roam the filesystem freely. The iMessage database is stored locally on the Mac but co-work needs folder access granted explicitly.
Unlock the full claims, asset map, scores, related transcripts, follow-up questions, and AI chat — shaped around your portfolio, watchlist, favorite speakers, and risks.
