Alerts that page humans, not noise.

How it works.

Every agent run produces structured signals: tokens, latency, retry count, tool error count, outcome class. Trefur learns the normal envelope per agent over a rolling window, then alerts when current behavior drifts outside it.

You configure routing per agent, per environment, per severity. Pick the channel — Slack, PagerDuty, webhook, email, Microsoft Teams. Pick the schedule and the escalation. Each alert includes a one-line summary and a deep link to the worst trace in the window.

Quiet hours, deduplication windows, acknowledgement, and snooze are first-class. The envelope learns each agent's normal cost, latency, retry, and error behavior over a rolling window, so a clean deploy settles into the new baseline instead of paging the whole team during warm-up.

What you can do with it.

An alert lands in Slack.

Routes to.