Agent Reliability

Agent alerts that name the cause

Detecting a failure and getting the right person to see it are two different problems. Tessary routes reliability signals to Slack, Sentry, Linear, PagerDuty, or a webhook, calibrated so the destination doesn't fill with noise. And the alert names the change that caused the regression, not just the score that moved.

Find what broke your agent
The gap

Detection outpaces delivery

In March 2026, PagerDuty added more than thirty AI-agent integration partners across eleven categories, most built to push raw signals like error spikes and latency jumps into an on-call rotation. Routing a raw score spike to PagerDuty just moves the mystery to whoever is on call: the page says a number moved, not which change moved it. A page worth waking up for carries the cause.

PagerDuty integration ecosystem ↗

30+

AI-agent integration partners PagerDuty added in March 2026

11

partner categories those integrations span
What Tessary Routes

Four alert types, five destinations

Each signal picks its own alert type and destination, so routing matches how the team already triages.

Digest

Rolls up signal activity over a set window. Built for a Slack channel someone skims each morning, not for immediate action.

Brief

A focused summary of one incident: the grader dimension that broke, the trend, and the change Tessary traced it to (a prompt edit, a tool update, or a model provider change that never touched your repo). Enough to act on without digging through raw traces.

Threshold alert

Fires when the confirmed failure rate for a signal crosses the rate you set. Tessary reads the trend across grader verdicts rather than any single score, so one imperfect grader misfiring does not page anyone. Meant for destinations like PagerDuty, where immediate action matters.

Pre-deploy risk report

When a pull request touches code linked to a known failure family (tool-call errors, hallucinated citations, refusal spikes), Tessary posts a risk report to that family's alert destination. The reviewer sees the warning in the same channel the production alert would land in.
Where It Lands

Point each signal at its destination

Choose per signal, not per project. A digest and a page can come from the same detection.

  1. Slack

    A channel someone checks periodically. Good for digests and briefs that don't need an immediate response.

  2. Sentry

    Keeps agent failures next to the error tracking the engineering team already watches.

  3. Linear

    Opens a ticket directly, useful when a failure needs to enter the existing backlog instead of interrupting someone.

  4. PagerDuty

    Pages the on-call rotation. Reserved for threshold alerts on failure families that should wake someone up.

  5. Webhook

    A generic endpoint for routing into any tool that isn't a first-class destination yet.
Precision modes

Tune sensitivity as trust builds

A threshold set too low pages someone for every agent quirk. Set too high, and the failure that mattered gets buried next to the noise. Tessary's precision modes handle the difference.

Discovery mode

For a signal you are still tuning

  • Surfaces everything the signal catches, false positives included
  • Useful for seeing the full range of what a new signal detects
  • Not meant to route to a paging destination yet
Tessary

Tracking mode

For a signal you trust

  • Alerts only on confirmed instances of the failure
  • Destination stops filling with quirks that are not real failures
  • Safe to route to PagerDuty once the signal has been proven out
The gap in one line
Your dashboards are green. Your agent is on fire.
Jamie MallersOneUptimeMarch 2026 analysis ↗
Before alertingAlready need traces flowing first? See agent observability setup.
FAQ

Questions about alert routing

Tessary routes agent reliability signals to Slack, Sentry, Linear, PagerDuty, or a generic webhook. You choose the destination per signal, so a digest can land in a Slack channel someone skims each morning while a threshold alert for a critical failure family pages the on-call rotation directly in PagerDuty. Whatever the destination, the alert carries the change Tessary attributed the regression to, so triage starts at the fix.
A digest is a rollup of signal activity over a set window, useful for a channel someone checks periodically. A brief is a focused summary of one incident, including the change behind it. A threshold alert fires when the confirmed failure rate crosses the rate you set, read as a trend across grader verdicts rather than a single score, and is meant for destinations like PagerDuty where immediate action matters. A fourth type, the pre-deploy risk report, warns a PR reviewer before the change ships.
Tessary's precision modes separate a new signal from a trusted one. Discovery mode surfaces everything a signal catches, including likely false positives, while you tune it. Tracking mode switches to alerting only on confirmed instances once you trust the signal, so the destination stops filling with quirks that aren't real failures.
You need traces landing in Tessary first, either through the OTLP receiver or by pulling a slice from Langfuse or Braintrust. Once traces are flowing and signals are detecting failures, alert routing is a separate configuration step: choose a destination and a threshold per signal.
Yes. When a pull request's diff is predicted to affect a known failure family, the risk report posts to the same destination as that family's production alert, so the reviewer and the on-call engineer look at the same evidence. It is the same attribution engine, run before the merge. Note the boundary: a pre-deploy check only covers changes that ship through a PR. A model provider update or an upstream agent's shift reintroduces a failure with no deploy at all, which is why production alerting keeps running.
Langfuse and Braintrust hold your traces and can alert you that a score moved. A Tessary alert also names what moved it, so on-call starts at the fix instead of at the investigation. Keep them as your trace store; Tessary pulls from both.
Keep exploring

Related pages
Get started

Get alerts that name the cause

Connect Slack, Sentry, Linear, or PagerDuty, with no credit card required, and the next regression arrives as an alert that says which change caused it.

Find what broke your agent