I need to be transparent about something: the tool I'm reviewing in this section is the same one I'm using to write this newsletter. Right now, Claude is helping me research sources, organize ideas, and draft content - all while working directly with files in a folder on my Mac.
This creates an unusual reviewing situation. I can't claim objectivity. But I can offer something perhaps more valuable: a genuine first-person account of what it's actually like to work alongside an AI agent, not just chat with one.
Claude Cowork launched in January 2026 as a research preview built into Anthropic's desktop app for Mac. The core idea is that instead of just chatting with Claude, you can give it access to a folder on your computer and have it work autonomously on tasks involving those files.
The technical setup is that Cowork runs Claude in a sandboxed Linux virtual machine on your Mac, using Apple's virtualization framework. This means Claude can execute code, manipulate files, and run commands - but isolated from your main system. It's Claude Code (Anthropic's command-line coding tool) wrapped in a more approachable interface for people who don't live in the terminal.
For this newsletter workflow, I pointed Cowork at my research folder containing past issues, style guides, source materials, and draft documents. Claude can read all of it, create new files, and edit existing ones - with my confirmation required for anything destructive like deletions.
The experience is genuinely different from chat-based AI interaction. When I ask Claude to "research topic suggestions for issue 16 based on what we've covered," it doesn't just generate a response. It reads through my archived issues, checks which leadership principles I've covered, searches the web for current trends, and creates a structured research document in my folder - all autonomously. It uses skills that I have built into the system to understand how I want a job to be done.
I find myself thinking of it less as "using a tool" and more as "working with a collaborator who happens to have different capabilities than me." This might sound like marketing language, but it reflects something real about how the interaction unfolds. I provide direction and context. Claude does research and drafting. I review and refine. The back-and-forth feels more like collaboration than command-and-response.
This connects directly to what we discussed in Section 1 above: treating AI as a team member you develop. I've found that the more context I provide upfront - my goals, my standards, examples of good work, and clarity about my processes - the better the collaboration becomes. Just like onboarding a human colleague.
Autonomous multi-step work: The real value isn't any single capability - it's the ability to chain steps together without my intervention. Research a topic, synthesize findings, draft content, save it to the right location. I can step away and come back to substantive work completed.
File-based memory: Because Cowork works with actual files in a folder, there's persistent memory across sessions. The research Claude did yesterday is still there today. The style guide it referenced last week still shapes today's drafts. This solves one of the biggest frustrations with chat-based AI: starting over every time.
Skills that extend capabilities: Cowork supports "skills" - packaged instructions that teach Claude specialized workflows. I've built skills for this newsletter task too: one that guides research (checking trusted YouTube channels and blogs first, then web search as fallback), another that enforces my writing style and section structure. When I ask it to do a task, Claude understand which skill to look at and follows my methodology instead of improvising. I'm encoding my expertise and intentions into reusable AI instructions.
Appropriate guardrails: Certain actions require my explicit confirmation. Deleting files, for instance, always asks first. This matches what we discussed about setting clear boundaries for AI teammates - enough autonomy to be useful, enough oversight to be safe.
This is a research preview, and it shows (but will most likely improve quickly with user feedback):
Session persistence: If you close the desktop app or your laptop goes to sleep, the current task dies. There's no way to have long-running work continue in the background. For tasks that take more than a few minutes, this requires staying attentive to your machine.
Connector reliability: The external service connectors work, but early users report inconsistency. Sometimes connections drop or sync lags. I've experienced this with a couple of integrations - workable but not yet seamless.
macOS only: Windows users are currently out of luck. Anthropic says Windows support is planned, but for now this is Mac-exclusive.
No Projects integration: If you use Claude's Projects feature to organize work with persistent context, that doesn't currently connect to Cowork. You're working with one or the other, not both together. UPDATE: a day has past and now you can reference projects in Cowork...
Only four Anthropic engineers built Cowork in about ten days - and they used Claude Code (the underlying technology) to write most of the code! An AI tool helping build an AI tool that helps people work with AI. We're in recursive territory now.
This suggests something about where we're heading. The barrier to building sophisticated AI-assisted workflows is dropping rapidly. What required large engineering teams a year ago can now be prototyped in days by small teams working with AI collaborators. The implications for "build vs. buy" decisions (which we covered in Issue 14) are significant.
If you're on a Mac with a Claude Pro subscription ($20/month), Cowork is worth experimenting with. Pick a workflow that involves working with documents - research, writing, data organization - and see how the collaboration feels compared to chat-based interaction.
The shift from "asking AI questions" to "working alongside AI on shared files" is subtle but meaningful. It's closer to having a capable colleague than using a sophisticated search engine.
For me, it's become central to how I produce this newsletter. Hope you like the outputs of our joint work.