What are scheduled AI agents? Routines, explained

A scheduled AI agent is an AI agent that runs on a schedule instead of waiting for you to prompt it: you describe a recurring task once in plain language — check these pages, triage these issues, summarize this inbox — pick a cadence, and the agent runs unattended from then on, using your tools and reporting back with the result. Think cron, but the thing on the schedule can read, reason, and write.

The pairing matters more than either half. AI chat is powerful but sits idle until you show up with a prompt. Schedulers are reliable but dumb — they can fire a script, not exercise judgment. Put an agent on the scheduler and you get a new category: recurring work that needs a bit of thinking, done without you.

What makes it an agent, not an automation

A classic automation — a Zap, a scenario, a workflow — is a fixed path: when trigger, do steps. It's excellent when the steps never change. A scheduled agent gets a goal, not a path. Each run it looks at what's actually there and works out the steps: which pages moved, which issues matter, what changed since last time, what's worth telling you. When the input shifts — a page redesign, a new format, an ambiguous case — there's no broken node, because there were never any nodes.

That's why the tasks that suit scheduled agents are the ones automation builders quietly give up on — the judgment-shaped ones:

• Monitoring with taste: competitor pricing, brand mentions, changelog watches — where "what changed" needs interpreting, not just detecting.
• Digests worth reading: a morning brief across your inbox, calendar, and news that says what matters, not everything that happened.
• Recurring research: keyword opportunities, lead lists, market scans — compiled fresh on a cadence.
• Triage: new GitHub issues, support tickets, or form submissions sorted and summarized before you sit down.

The anatomy of a routine

In Clourou we call one scheduled task a routine. A routine has a prompt (the plain-language brief), a schedule (real cron, any timezone, down to every 15 minutes), tools (any MCP server you attach — Gmail, Notion, GitHub, or your own), and a delivery setting (results land in-app, and by email if you opt in). Each run gets a real Linux sandbox, files persist between runs, and every routine keeps an editable memory file — so a digest doesn't repeat yesterday's items and your corrections stick.

The unglamorous parts are what make it dependable: failed runs retry and then alert you by email, every run keeps a full tool-call trace you can audit, and a spend cap you set bounds what an unattended schedule can ever cost.

Where they fit next to what you already use

Scheduled agents don't replace workflow tools — they cover the other half of the map. If your task is a fixed pipeline with known steps, a deterministic builder is cheaper and more auditable; we say so plainly in our comparisons. If your task needs reading, comparing, and writing — the things you'd otherwise do by hand on a recurring basis — that's agent territory. The dividing line is simple: can you write the steps down in advance? If yes, automate them. If no, schedule an agent.

See one run

The fastest way to get the idea is to watch one work. Browse the routine templates — a daily competitor pricing watch, a morning inbox briefing, a weekly GitHub release digest — each with its exact prompt and a sample of the email it sends. Copy one, and it's live on a schedule in a click.

For the wider pattern — what belongs on an agent's schedule and how MCP makes the tool side work — read 12 things to put on an AI agent's schedule and the MCP-native automation playbook.