Anthropic expanded Claude Cowork from the desktop —where it launched in January— to web and mobile, starting with the Max plan: sessions are hosted in Anthropic's cloud and keep running agentic loops and scheduled tasks even when no device is turned on. An analysis of 1.2 million sessions (May 2-31) from more than 600,000 organizations shows that only 8.7% of usage is software development; 33.4% is running business processes.
From the desktop to your pocket
Picture handing a long task to a brilliant intern, closing your laptop, catching the subway home — and by the time you walk in the door, the work is done. Not because the intern stayed late at the office, but because they never needed you sitting there. That's exactly what Anthropic just turned into a product: on July 7 it expanded Claude Cowork —which had launched in January as a desktop app— to the web and mobile, starting with the Max plan. In doing so, it moved the agent's work to a place where your device stops being the bottleneck. The session no longer lives on your machine. It lives in the cloud.
Working with the laptop closed
The technical detail is the one that changes everything. Cowork's remote sessions are hosted on Anthropic's infrastructure, saved to your account, and keep running multi-step agentic loops —and scheduled tasks— even when you don't have a single device turned on. In plain terms: you kick off a task from your phone while standing in the checkout line, put the phone away, and the agent keeps working on its own thread while you do something else. For years, AI assistants were like a phone call: they only existed while you held them in your hand. Cowork turns them into something more like an employee with their own desk, one that doesn't shut off when you leave. That asynchrony —being able to drop the task and trust it won't die— is the real leap, far more than the fact that there's now an app on your phone. The desktop, mind you, is still home to deep work with local files and a browser; the web and mobile are the hands that reach it from the outside.
The data point that dismantles the "agent = coder" myth
And then Anthropic dropped the figure that should be framed on the wall of everyone building AI products. It analyzed 1.2 million anonymized sessions —collected between May 2 and May 31— from more than 600,000 organizations, and the breakdown is devastating: 33.4% of usage was running business processes —finance, HR, admin—, 16.4% content creation, and only 8.7% software development. Read that again. The "code agent wars" that have filled headlines for a year turn out to describe fewer than one in ten real uses. The vast majority of people don't use these agents to write code: they use them for the work around the work — reconciling a spreadsheet, prepping a report, cross-referencing data between tools, drafting and sending. The office, not the IDE.
Why this reshuffles who products are aimed at
The underlying message for the whole industry is uncomfortable and freeing at the same time. Uncomfortable because so many AI tools have sold themselves as "coding copilots," chasing that 8.7% with hacker marketing. Freeing because the real market —the remaining 90%— is people who don't know how to code and don't want to, and who just need a boring multi-step task to run itself while they get on with their own thing. The value isn't in writing the perfect function; it's in running the business process end to end without friction. Whoever gets that stops building for programmers and starts building for the rest of the office, which is where nearly everyone actually is.
How this connects to NeuralOS
This news validates, almost point for point, a bet NeuralOS already has in its engine: the agent's work shouldn't depend on you staring at the screen. Our Automations engine is resumable and ships with durable waits — meaning a flow can wait hours or days for an external condition, survive you closing the session, and pick up exactly where it left off. That's the same "asynchronous office agent" pattern Cowork just popularized: drop the task and trust it keeps going. We'll tell you straight, no smoke: NeuralOS isn't a Cowork competitor and doesn't aim to be, and some of our pieces (the brain, the graph RAG) are still vision in the UI with a backend on the roadmap, not magic that's ready today. The honest lesson we take from those 1.2 million sessions is one of strategy, not features: the agent that really matters to most people isn't the one that codes, it's the one that handles the business process and doesn't force you to stay on call. Building for that person —the one who wants results, not code— is the path we're taking.