The routine that runs itself.
A developer talk making the rounds this week described AI work running on a set schedule with no one typing a fresh instruction each time. Anthropic's own team, the talk said, now has AI running twenty to thirty of these scheduled jobs a day against its own codebase, cleaning up loose ends while engineers get on with other things. The framing is new. Work that repeats shouldn't need to be asked for twice — that part isn't.
What was actually said
The talk, given on stage at Startup School, was from Anthropic's Boris Cherny, walking through how Claude Code has changed under the hood. Buried in the technical detail was a plain distinction between three ways AI work can run: a single big job you kick off once and let run for days, a job that repeats on a local schedule, and the same repeating job running in the cloud so nobody needs a laptop open for it. Anthropic said it now runs a stack of these repeating jobs daily against its own engineering work — the same dead-code tidying and duplicate-cleanup admin that piles up in any team, just handled on a schedule rather than remembered by a person.
None of that is a new AI capability arriving for you to adopt. It's a naming of something that was already possible, and a public example of a team using it on their own repeat admin instead of one-off requests.
Where this actually bites
The bottleneck it points at isn't exotic. It's the admin tax: inbox triage, the same filing done every Friday, the same follow-up chased every Monday. None of it needs a person's judgement each time. It needs a person to remember to start it. That's precisely the step that gets skipped when the week gets busy.
Most small firms run this admin the same way: as a favour to future-you, done when there's a spare hour. Some weeks that hour doesn't arrive. The work doesn't stop being necessary. It just stops happening on time.
What still needs a person
Two things worth being straight about.
A schedule doesn't know what "done properly" means for your business. Anthropic's own routines are checked against test suites built for exactly that purpose. Your equivalent check is your brand kit and your own approval step. Without them, a scheduled job just repeats whatever it was told, including a mistake, on time, every time.
Running more often isn't the same as running unwatched. A job that repeats daily needs a log more, not less, than a one-off request. If nobody can see what a recurring job actually did last Tuesday, more automation just means more unread output.
Where WorkMate OS fits this
Recurring admin already runs this way inside WorkMate OS, and there's nothing to set up to get it. Filing, inbox triage and routine chasing sit inside your crew's normal working day, the same as any other job. There's no separate schedule to write, because the crew's day already includes the jobs that repeat.
The check that stops a schedule going wrong quietly is the same one that stops any WorkMate job going wrong: outward-facing work goes draft, then review, then send, and you keep the final word. And the one shared Brain means a recurring job is checked against your actual brand kit and terms, not a generic template repeated on autopilot.
The other half is the activity log: you see what ran, when, and what it produced, so a job running daily is no harder to keep an eye on than one you ran once yourself. That's also the honest answer to the cost question this kind of story raises. Everyday scheduled admin sits inside your flat monthly plan, with no per-task meter. We wrote the meter question up properly in the meter problem.
On the news itself: there's nothing to install or configure to get this. If a supplier is describing a routine you need to set up yourself, they're handing you the maintenance job a proper AI teammate should be doing without being asked twice.
If you do nothing else this week
Write down the three admin jobs you do on the same rough schedule every week: filing, chasing, triage. That list is the actual test for whether an AI teammate is doing its job. Not whether it can have a clever conversation. Whether those three things stop needing you to remember them.
