Guide

Scheduling AI agent routines: cron patterns that stay cheap

A routine is what turns an agent from “something you remember to ask” into a job that runs itself. The hard part isn’t the cron line — it’s picking when to wake, and what to do when there’s nothing worth doing.

1. Pick the trigger before the schedule

Ask first: does something happen that should wake the agent (a PR opened, a message posted, a form submitted)? If yes, use an event trigger or webhook. GitHub’s own docs put it plainly: webhooks deliver data as it happens and take less effort and fewer resources than polling an API. Save cron for work that is genuinely time-based — digests, weekly reviews, “check the queue each morning.”

• Event trigger — PR opened, CI finished, Slack message, incoming webhook. Runs only when there’s input.
• Schedule — daily brief, weekday publishing slot, end-of-week summary.
• Both — many tools let one routine have several triggers; it runs when any of them fires.

2. Cron patterns worth copying

Standard five-field cron (minute hour day-of-month month day-of-week):

• 7 8 * * 1-5 — 08:07 on weekdays. The classic “morning brief” slot.
• 37 17 * * 5 — 17:37 Friday. Weekly wrap-up or review.
• 17 9-17/2 * * 1-5 — every two hours during a weekday working window.
• 23 7 1 * * — first of the month, for monthly reports or cleanups.

Check the timezone: Cloudflare Cron Triggers run on UTC, and GitHub Actions schedules default to UTC unless you set one. Cloudflare also numbers weekdays 1 = Sunday, unlike most cron systems — use MON–FRI names to avoid surprises.

3. Avoid the top of the hour

Everyone schedules at :00. GitHub notes that scheduled workflows can be delayed during high load, which includes the start of every hour, and recommends a different minute. Cursor’s scheduled triggers may also start late (never early). Pick an odd minute like :07 or :43, and don’t write routines that break if they run a few minutes late.

4. Be quiet when nothing changed

The most useful rule for a scheduled agent: if there’s nothing new, do nothing and say nothing. A daily “no updates today” message trains people to ignore the bot.

• Check cheaply first — look at a timestamp, count, or hash before invoking a model at all.
• Write a quality bar into the prompt — spell out when to comment, open a PR, publish, or do nothing.
• Keep a small memory — store what was already handled so the next run skips it.
• Exit silently — no summary, no ping, no deploy when the answer is “nothing.”

5. Keep token and usage cost down

• Run less often — hourly is 24× as many wakes as daily. Most digests don’t need to be fresher than once a morning.
• Narrow the input — pass only what changed since the last run, not the whole inbox or repo.
• Keep the prompt prefix stable — providers such as xAI cache repeated leading messages automatically and bill cached tokens at a reduced rate, so put fixed instructions first and variable data last.
• Start simple — a single well-scoped call on a schedule often beats a multi-step agent loop. Add agentic steps only when the task needs them.
• One job per routine — a narrow job has smaller context and is easier to tell whether it did anything.

6. A quick checklist

• Is there an event I could trigger on instead of polling?
• Is the schedule in the timezone I think it is?
• Is it off the top of the hour?
• Does it exit silently when there’s nothing to do?
• Does it remember what it already handled?
• Could it run half as often without anyone noticing?

Useful public resources

New to narrow bots? Start with What is a one-job Grok Bot?