Alert Rules
Alert rules watch your metrics and notify you the moment a threshold is breached, so you don't have to stare at a dashboard.
Anatomy of a rule
A rule combines a metric, an operator, a threshold, and a rolling window. The window smooths out one-off blips so the alert that reaches you is one worth acting on.
Metrics
| Metric | Catches |
|---|---|
error_rate | A deploy or upstream change breaking runs. |
success_rate | The inverse view — alert when it drops below a floor. |
avg_latency_ms | A slow tool or degraded model provider. |
failed_runs | Absolute count of failures — useful at high volume. |
cost_usd | A retry loop or ballooning prompt burning budget. |
Operators and windows
Use gt (greater than) or lt (less than) against your threshold, evaluated over the rolling window you choose. For example, error_rate gt 5 over 15 minutes fires when more than 5% of runs in the last 15 minutes failed.
Deduplication and re-arming
AgentPulse deduplicates open alerts — you get one notification per breach, not one per minute. The rule re-arms only after the condition resolves, so you are told once when it breaks and once when it is healthy again.
Delivery channels
Attach one or more channels to a rule — Slack, a generic HTTP webhook, or email. See Notifications to configure them. Delivery is a no-op for any channel that is not configured.