The dream is an assistant that does recurring work without you babysitting it — and Cowork will actually do that, if you build the loop instead of re-running the task by hand every day. The trick is designing systems that (a) fire on their own and (b) don’t pay full price every time they run.
Scheduled tasks: the thing most people don’t know exists
Anything you’d want “every morning” or “each Monday” can be a scheduled task: a daily briefing of your inbox, a weekly digest of what changed in a project, a monitor that pings you only when some condition trips. You describe it once and it runs on a cron-like schedule, delivering the result back to you. The mental shift is from “I ask for this every day” to “this shows up whether I ask or not” — same output, none of the recurring effort.
Live artifacts: build once, refresh forever
For anything you’ll want to look at again — a status board, a metrics page, a tracker — a live artifact beats regenerating a report. It’s a small self-contained page that pulls fresh data from your connectors each time you open it, rather than re-running an expensive generation. One build, unlimited refreshes, and the reads are cached so you’re not paying to rebuild the whole thing every glance. A markdown table you’ll ask for again next week is a report; a page you reopen is an artifact. Build the second one.
The AI tool stack actually worth paying for
One email a week. The models, tools and moves that matter, stripped of hype and filtered so you don’t have to drink from the firehose. Free, and you can bail anytime.
Design loops that don’t redo work
This is where cheap loops are won or lost. A naive daily job re-processes everything from scratch; a smart one processes only what changed. Have the loop remember state — the last item it handled, the last timestamp it saw — and act only on the delta. Summarizing “new emails since yesterday” should read yesterday’s, not your entire inbox. Incremental beats total every time, and it’s the difference between a loop that costs pennies and one that costs a fortune to run daily.
Tier the model to the job
Your recurring loop almost never needs your most powerful model. Routine classification, filtering, and “did anything important happen” checks run fine on a fast, cheap model; you escalate to the heavy one only for the rare step that genuinely needs deep reasoning. Match the model to the difficulty of each step and a daily automation stays comfortably affordable. Set-and-forget shouldn’t mean set-and-go-broke — a well-built loop fires on schedule, touches only new work, and uses the cheapest tool that does the job.