How often should your agent run?

Measure how fast the answer loses value

Start by estimating the shelf life of the output: how long after a run could you still act on it and get the same result? A support escalation may be worthless in twenty minutes. A competitor summary may hold up for a week. A monthly market read is fine a few days late. Tie the cadence to that decay, not to how often the button can be pressed.

How to judge it: look at your last run and ask whether the same output, delivered a day later, would have changed any decision. If the honest answer is no, you are refreshing faster than the information decays.

The failure here is quiet: an hourly “competitive watch” that nobody revisits until the next strategy session, generating freshness no one ever spends. The trade-off runs the other way too—slow the cadence past the decay point and you act on stale facts—so the goal is to sit just inside the shelf life, not far below it.

Check how often the source actually changes

Before choosing a frequency, sample the source: have the agent watch it for two weeks and simply log when it meaningfully changes. Checking a page many times more often than it moves does not surface more—it files the same report again and again. The rough rule from signal sampling is to check a little faster than the thing changes, not orders of magnitude faster.

How to judge it: divide runs by meaningful changes. If a hundred runs surfaced one real change, ninety-nine were noise and the cadence is far too high.

The failure is a source edited monthly polled every hour: hundreds of identical outputs with the one real change buried inside them. The trade-off is that a slower schedule can miss a change that happens between checks—so pair the slow cadence with an event trigger for the rare urgent case rather than speeding the whole thing up.

Respect how much you can actually read

An agent can out-produce any reader. If the next step—you, a teammate, or another workflow—can absorb ten items a day, producing fifty does not create more value; it creates a backlog and, worse, a habit of ignoring the feed. Match output volume to the capacity of whoever reads it next.

How to judge it: track produced versus opened. If most outputs go unread, the frequency is tuned for the agent, not for you.

The failure has a name outside software—alert fatigue—where a flood of low-value notices trains people to dismiss everything, including the one notice that mattered. The trade-off is immediacy: batching a stream into one daily digest costs you same-minute awareness but buys back the attention that makes any of it worth reading.

Start at the lowest useful cadence

Set the first schedule deliberately slow and let evidence talk you into faster—adding frequency later is a two-second change, while unwinding a noisy habit means admitting the whole feed became wallpaper. Use this ladder as a default:

• Weekly for strategic research and slow-moving competitive reads.
• Daily for operational summaries and morning briefs.
• Hourly only when a delayed reaction has a clear, nameable cost.
• Event-triggered for urgent, irregular changes that do not fit any fixed clock.

How to judge it: after the first week you should feel a mild wish that it were a little faster—not relief that it finally went quiet.

The failure is setting it hourly “to be safe” on day one and never revisiting, so noise becomes permanent. The trade-off is that starting slow risks a slightly delayed first catch; that cost is almost always smaller than the chronic tax of a feed you have learned to skip.

Re-check after two full cycles

The right frequency is discovered, not guessed. After two complete cycles, ask three questions: were any decisions delayed waiting for a run, did reports keep repeating the same information, and did outputs go unread? Each symptom points one direction—too slow, too fast, too much—and you move one level to match.

How to judge it: one clear symptom, one level of change—monthly to weekly, weekly to daily, or a fixed clock to an event trigger.

The failure is impatience dressed as diligence: changing the cadence and the scope in the same week, so when things improve you cannot tell which lever did it. The trade-off is that waiting two cycles delays the fix—but it is the difference between tuning on a signal and tuning on a mood.