This month Anthropic put its managed agents on a timer. You can now hand an agent a schedule, and each time it fires it wakes up, does a defined job, and goes quiet again, with no scheduling system for you to build or maintain. Big firms are already using it for unglamorous things: Rakuten runs a spreadsheet-to-report job on a weekly clock. For a small business, the question isn't whether this is impressive. It's which of your jobs wants doing on their own, and which absolutely should not. Here's the line I draw for clients, and the three jobs I'd put on the night shift first.
For most of the last two years, “AI agent” has meant a thing you sit in front of. You open a window, you ask, it answers, you act on the answer. Useful, but it still needs you in the chair. The change worth noticing this month is small and structural: you can now give one of these agents a schedule. It clocks on by itself, does the job you defined, writes down what it found, and clocks off. Nobody opens a window. Nobody presses go.
Anthropic shipped scheduled deployments for its managed agents in early June, and the framing in their own announcement is the honest one: use it for “recurring work like a nightly data sync, a weekly compliance scan, or a daily digest.” There's no clever scheduler for you to host — you give the agent a time, and the platform handles waking it up. Rakuten, the Japanese commerce and finance group, already runs spreadsheet analysis on a timer that produces a report on a weekly or monthly clock. One sales-tools company said it deleted the scheduling system it had built in-house and replaced it with this. That's the tell. When sophisticated teams throw away their own plumbing, the plumbing just became a commodity.
Why this matters more for a small business than a big one
A large company has people whose entire job is to run things on a schedule. A small business doesn't. The Monday-morning report gets done because someone gives up the first hour of their Monday to do it. The weekly check on which invoices have slipped gets done when somebody remembers. The “has anything broken over the weekend” sweep happens when you, the owner, open your laptop on Sunday night with a coffee and a vague sense of dread.
Those jobs share a shape. They're recurring, they're rules-based, and they mostly produce a short answer: here's what changed, here's what needs a human. That shape is exactly what a scheduled agent is good at. It's not doing the judgement. It's doing the gathering — the part that eats your time and gives nothing back — and handing you a tidy answer to react to.
I'll be straight about my own bias here: I run my consultancy on this pattern. A handful of routines wake up on their own through the week, read what they're pointed at, and leave me a digest to glance at. I didn't build them because they're clever. I built them because the alternative was me doing the same dull sweep by hand, slightly worse, slightly later, every single week. The night shift doesn't get tired and doesn't forget on a busy Tuesday.
The line I draw: gather on a timer, decide in daylight
Here's the rule, and it's the whole article really. A scheduled agent should gather, sort, and flag. A human should decide and send. The moment a job crosses from “tell me what's true” into “do the irreversible thing” — pay the supplier, email the customer, change the price — it comes off the timer and onto a person's desk, even if the agent has already drafted the action.
Put plainly, three questions decide whether a job belongs on the night shift:
- Does it repeat on a clock? If it happens every Monday, or every month-end, or every night, it's a candidate. One-offs aren't.
- Is the output something I react to, not something the world feels? A digest in your inbox is safe. A payment leaving your account is not. The first can run unattended; the second needs your thumb on it.
- Would I notice if it quietly went wrong? If the answer is no, you're not ready to automate it yet. You're ready to define what “right” looks like first. Automating a job you can't check is how you end up trusting a number that drifted three weeks ago.
What I would not put on a timer yet
Anything that talks to a customer in your name. The technology can draft the reply beautifully; it cannot yet own the consequence of sending it at 3am with no one watching. A property tech client of mine has an agent that drafts responses to overnight enquiries. But it queues them for a human to glance at and release in the morning. The gathering ran unattended. The sending didn't. That single boundary is the difference between a helpful overnight worker and a liability you find out about from an angry email.
I'd also leave off anything where the rules change often. A scheduled agent is brilliant at a stable job and brittle at a moving one. If the task is “do exactly this every week,” automate it. If the task is “use your judgement about a situation that's different every time,” that's still yours.
The cost-benefit
The benefit isn't that you save a fortune. It's that you reclaim the specific, recurring hours you currently spend being a human cron job — the gathering, the checking, the “let me just pull that together” tax on every week. For most small business owners that's somewhere between two and five hours a week, and they're not your good hours. They're the admin you do when you should be selling or thinking.
The cost is the discipline of setting it up properly: writing down what the job is, naming what good output looks like, and keeping a human on anything irreversible. That's an afternoon, not a project. And the new scheduling tools remove the part that used to make this a developer job. There's no server to run, no timer to maintain.
So the question I'd sit with this week is a small one. Look at your calendar's repeating dread — the report, the sweep, the chase-up you do every week because no one else will. One of those almost certainly wants doing on a timer, by something that never has a busy Tuesday. Start with that one. Keep your thumb on anything that spends money or speaks to a customer. Let the rest run.