Alerts that page humans, not noise.
Agents fail differently than services. They retry quietly. They spend money quickly. They produce wrong answers without crashing. Trefur is the oncall surface that catches those failure modes — with the trace already pulled up.
What you are trying to do.
- →Keep agents reliable in production without owning every product team's framework.
- →Page humans only when something actually needs a human, with the trace already open.
- →Make the on-call runbook for AI agents look like the runbook for any other service.
- →Catch slow-burn regressions — retry creep, cost creep, latency creep — before they become incidents.
- →Hand off cleanly to product engineers with the failing trace, not a one-line log.
What hurts today.
Token-count alerts are useless
A spike in token usage tells you nothing about what failed. Was it a retry storm? A model swap? A long user input? You cannot triage without a trace.
Agents fail loud and unhelpful
The error says JSON parse failed. The next line says retry succeeded. The user got a wrong answer anyway. You need to see the decision the agent made, not the last line in stderr.
Oncall gets paged by the wrong thing
Every minor LLM hiccup turns into a Slack alert at 3am. After a week you mute the channel. Then the real incident slides past unnoticed.
Postmortems take forever
Reconstructing what an agent ran during an incident means joining four log streams across two services. By the time you understand it, the model has been swapped.
What Trefur gives you.
Alerts with traces attached
Every alert links straight to the worst trace in the window. Open Trefur, see the failing step, hand off with one URL.
Envelope-based detection
Alert on retry-rate spikes, cost-per-run drift, p95 latency walks, error-rate envelopes. Not on absolute token counts.
Standard routing
Slack, PagerDuty, Teams, webhook, email. Per agent, per environment, per severity. Quiet hours, deduplication, acknowledgement.
Search by the thing oncall has
Trace ID from the support ticket. Customer email. Agent name. Error class. Time window. Find the run in seconds.
Replay a failing run
See the exact prompt the model saw, the exact tool response it got, the exact retry it tried. Hand it to the product engineer without needing them to repro.
Quiet by default
Trefur learns each agent's normal cost, latency, and retry envelope. Alerts fire on drift, not on the first noisy minute.
A day in the life with Trefur.
- 02:14
PagerDuty pages — retry rate on the refund agent doubled over a 10-minute window. The alert links a trace where a tool returned 503 four times before the agent gave up.
- 02:17
You confirm the tool service is degraded, ack the page, and the agent's circuit breaker has already paused new runs. Two minutes of triage, not twenty.
- 08:30
Morning standup. You share the dashboard: 14 incidents last week, 11 with linked traces, average MTTR down 38% since rolling out Trefur.
- 11:00
You write a new alert: any agent whose cost-per-run drifts more than 35% over a 24h baseline pages the owning team in Slack with a top-5 worst traces list.
- 15:45
Postmortem on yesterday's outage. The trace shows the agent retried a deprecated tool endpoint twice. Action items write themselves.
Make agent oncall look like normal oncall.
Free tier to start. Team and Enterprise add retention, SAML, and quiet-hours routing.