Most of us still think of AI as something we open, ask a question, and close. Type a prompt, get an answer, move on. Anthropic's Claude Cowork, the version of Claude built for non-coding office work rather than software development, just made that mental model outdated. It can now run tasks on a schedule, in the cloud, and keep working even after you have shut your laptop and gone home.
If you have not looked at Cowork before, here is the short version. It is Claude's agentic mode for knowledge work: instead of chatting one message at a time, you hand it a real job such as sorting a messy downloads folder, compiling research from several files, or drafting a report, and it plans the steps and carries them out with access to your files, Gmail, Drive, Slack and other connected tools. It is available on Pro, Max, Team and Enterprise plans, so it sits behind a paid subscription, not the free tier most people first try Claude on.
What actually changed
Two updates matter here, and both are about Cowork acting without you watching it.
1. Scheduled and recurring tasks now run in the cloud
You can now save a task's instructions once and set how often it repeats. Anthropic's own documentation confirms these scheduled tasks run remotely: your computer can be asleep, or the Claude desktop app can be fully closed, and the task still fires on its cadence. A task can search Slack, pull from your files, do web research, or generate a report, using whatever connectors you have already set up.
In practical terms, this turns Cowork from "a tool I use when I remember to" into something closer to a standing instruction to a junior colleague. Set it up once on a Friday afternoon, and it is still running every Monday morning whether or not you are at your desk.
2. Claude can now act in the browser without asking permission every step
Claude in Chrome, the extension that lets Claude click, fill and navigate a web page on your behalf, is now generally available across every paid Claude plan. The newer part is that Claude can take a sequence of actions in the browser on its own, rather than pausing for your approval before each click. Anthropic says a safety classifier checks every action before it runs, to catch anything that does not match what you actually asked for.
An AI that only answers when asked is a search engine with better manners. An AI that can be told once and keep working on a schedule, inside your actual tools, is a different kind of colleague. The judgment required to use it well is different too.
Where this is genuinely useful
Think about the repetitive parts of your week that do not need your judgment, only your consistency.
That last instruction matters as much as the task itself. A recurring task should end with something you check, not something that goes out on its own. Weekly status summaries, a Monday morning digest of overnight emails, a standing research pull on a topic you track for clients: these are exactly the tasks that are useful every single week and boring enough that nobody minds an AI doing them first.
Where to be careful
Handing a tool ongoing access to your inbox, files and chat tools is a bigger decision than asking it one question. Before you set up a recurring task, be clear about three things: what data it can see, what it is allowed to do without asking you first, and who else might be affected if it gets something wrong. A scheduled task that only reads and summarises is low risk. One that sends emails or messages on your behalf needs a human checkpoint before anything leaves your name.
This is also not a tool for anything involving client financial data, personal information or unpublished numbers unless your organisation has specifically approved it for that use. Treat a recurring AI task the same way you would treat delegating to a new team member: start with something low stakes, watch what it produces for a few cycles, and only widen its scope once you trust the output.
What to actually do with this
You do not need to hand over your whole week to an AI agent to benefit from this update. Pick one task you currently do the same way every week, one that mostly involves pulling information together rather than deciding something, and try setting it up as a recurring Cowork task. Check the output for two or three cycles before you rely on it. If it holds up, you have quietly reclaimed an hour a week. If it does not, you have learned exactly where the tool's judgment runs out, which is useful information either way.
The bigger shift is less about this one feature and more about what it signals. AI tools are moving from things you operate to things you delegate to, and that changes what fluency with them actually requires: not just writing a good prompt, but knowing what to hand over, what to keep checking, and what never to automate at all. That is the kind of judgment the AiM AI Fluency Test is built to measure, across eighteen real tasks in about twelve minutes, no login required.
Try one recurring task this week on something genuinely repetitive. Watch it for a cycle or two before you trust it. That is a more useful habit than reading every new feature announcement that comes out of Anthropic, OpenAI or Google this year, and there will be plenty more of them.