Scheduling tasks in Hermes: from a simple reminder to real automation

Everything so far has been reactive: you ask, the agent does. The real shift comes when the agent starts doing things without being asked — pulling a summary before you wake up, refreshing its own knowledge base overnight, or reminding you about something weeks from now. That’s what a scheduler is for. This article sets up Hermes cron scheduled tasks, starting with a reminder so simple anyone can follow it, then showing how the same mechanism powers real automation.

Why schedule an agent at all?

An agent that only acts when you talk to it is a faster way to do manual work — useful, but you’re still the one pressing the button every time. A scheduled agent is different: it becomes operational. It can run a task at a set time, on a repeating interval, or once at a future moment, and hand you the result without you having to remember to start it.

The best scheduled jobs are small and have a clear output: a reminder, a daily summary, a maintenance task like rebuilding an index. They’re not “think forever” open-ended missions — they’re specific, bounded jobs that run reliably. Get comfortable with one simple job first, and the pattern extends to everything else.

Your first job: a library-book reminder

Here’s the simplest possible useful example. Your library books are due back in three weeks and you’ll certainly forget. Instead of a sticky note, hand it to your agent:

/cron add 3w "Remind me to return the library books"

That’s the whole thing. In three weeks, the agent fires this job and delivers the reminder. Because it’s a one-off, it cleans itself up after running — no leftover job cluttering your schedule. You can also just ask in plain chat: tell Hermes “remind me in three weeks to return the library books” and it sets up the same job for you.

Trivial as it looks, this is the entire mechanism in miniature: a time, an instruction, and a delivered result. Everything more advanced is a variation on it.

The scheduling formats

You don’t need to learn cryptic cron expressions. Hermes accepts plain, human-friendly schedules:

• A duration from now — 30m, 4h, 3w. Good for one-off reminders: “in 30 minutes”, “in three weeks”.
• A recurring interval — "every 2h", "every 1h". The job repeats on that cadence until you remove it.
• A daily time — a plain “every morning at 9am” style schedule for things you want at a fixed hour.
• Classic cron expressions — the traditional five-field form (like 0 9 * * 1-5 for weekdays at 9am) is still there if you want precise control.

The rule of thumb: use a duration for one-off reminders, and a recurring interval or daily time for anything that should keep happening. You can add, list, and remove jobs the same way — through a chat command or by simply asking the agent.

The one rule: self-contained prompts

This is the single most important thing to understand, and it trips up almost everyone at first. Each scheduled job runs in a fresh session. It does not have your current conversation, it doesn’t remember what you were just discussing, and it can’t see the context you have in front of you right now.

So the job’s instruction must be completely self-contained — everything the agent needs to do the task has to be inside the prompt itself. “Summarise that article I mentioned” will fail, because the scheduled session has no idea what “that article” was. “Fetch the top three AI stories from Hacker News and summarise each in two sentences” works, because it stands entirely on its own.

Write every scheduled prompt as if the agent is seeing it cold, with no memory of you — because that’s exactly the situation. Name the sources, spell out the steps, state the desired output. A self-contained prompt is the difference between a job that works every time and one that quietly does nothing.

Getting the result: delivery

A scheduled job that runs but has nowhere to send its result isn’t much use. Hermes delivers cron output to wherever you point it — commonly a messaging channel like Telegram, so the reminder or summary lands on your phone. The delivered message is wrapped so you can tell it came from a scheduled task rather than a live conversation.

If you haven’t set up Telegram yet, that’s covered in the previous article on the Telegram gateway — once connected, scheduled results land straight in your Telegram chat.

Connecting a delivery channel like Telegram is its own small setup, and it’s what makes scheduled jobs genuinely useful on the go — your agent reaching out to you, rather than you always reaching for it. That gateway setup is covered separately in the series; once it’s in place, any scheduled job can deliver straight to your pocket.

From reminders to real automation

Once the mechanism clicks, the same tool does far more than reminders. A few examples that follow directly from earlier parts of this series:

• Refresh the vault index nightly. The vault index needs regenerating as your notes change. A daily early-morning job rebuilds it automatically, so it’s always current — exactly the kind of small, self-contained maintenance cron is built for.
• A morning briefing. “Every morning at 9am, fetch the top AI headlines and send me a two-line summary of each” — self-contained, recurring, delivered to Telegram.
• Periodic checks. Watch a page, a feed, or a status on an interval and report only when something changes.

One safety detail worth knowing: a scheduled job can’t create more scheduled jobs. Hermes deliberately disables that from inside a cron run, so a job can’t spawn a runaway chain of new jobs. It’s a guardrail that keeps automation from multiplying out of control.

Frequently asked questions

Do I need to learn cron syntax?

No. Plain durations like 3w and phrases like “every morning at 9am” cover most needs, and you can set jobs up just by asking the agent in chat. The classic five-field cron expressions are available if you want precise control, but they’re optional.

Why did my scheduled job do nothing?

Almost always because the prompt wasn’t self-contained. A scheduled job runs in a fresh session with no access to your current chat, so any instruction that refers to “that thing we discussed” has nothing to work with. Rewrite it so everything it needs is in the prompt itself.

What’s the difference between a one-off and a recurring job?

A one-off runs once at a future time and then removes itself — ideal for reminders. A recurring job runs on its interval or daily time and keeps going until you delete it — ideal for briefings, checks, and maintenance.

Where do the results go?

Wherever you configure delivery — commonly a Telegram channel so results reach your phone. The output is marked as coming from a scheduled task, so you can tell it apart from a live reply.

What’s next

Your agent can now act on its own schedule — from a three-week library reminder to a nightly index rebuild to a morning briefing. The two things to remember: use plain schedules rather than cryptic syntax, and always write self-contained prompts, because each job runs cold.

Next up: your agent now acts on its own schedule. A later article gives it the ability to read any website reliably, so your scheduled jobs can pull from the web without getting stuck on anti-bot walls. This article is part of the Agentic AI series. If you are just getting started, begin with the Windows 11 install guide.