Alerts for AI agents
Hear about it first. Not from a customer.
Alert rules watch every run for failures, error rates, slow runs, silence, your own events and budgets, and tell the right people in the app, by email or on a webhook.
The problem
Agents fail quietly.
A source starts blocking the agent and it writes half a report. A scheduled agent stops running and nothing raises an error, because nothing ran. The agent finishes, says something reasonable, and nobody notices for a week. An alert is how a problem reaches a person on the day it starts.
- It failed. A run ended with an error, once or more often than it should.
- It slowed down. A run that suddenly takes much longer often means a loop or a slow dependency.
- It went quiet. An agent that should have run by now and did not.
Conditions
Match the alert to how the agent fails.
A rule watches one agent, or every agent in the workspace when you leave the agent blank. Pick the condition that fits: an agent that must never fail gets an alert on any failure; a busy agent where one failure is noise gets one on its error rate.
- Any run fails. Fires when a run ends as failed. No threshold.
- Error rate exceeds threshold. The share of failed runs among the most recent ones, 10% of the last 20 by default.
- Run duration exceeds threshold. A run takes longer than the limit you set.
- Agent has no activity. No run started within a window, 24 hours by default. An agent that has never run counts too.
- Custom event type observed. Your code emits an event, such as
order.failed, and the rule fires on it. - Intelligence finding escalates. A built-in analyst raises its finding about an agent to warn or to critical. It fires as soon as the finding is written.
- Monthly budget exceeded. Estimated month-to-date spend crosses the budget for an agent or the whole workspace.
Channels
Sent where someone will see it.
Every alert lands in the app, on the Notifications page, with an unread badge on the user menu. Add email for up to ten recipients, a webhook for your own systems, or both.
A webhook is a JSON POST with the rule, its condition and what triggered it: the run and its error, the observed error rate, or the spend against the budget. A delivery that fails is retried twice, a few minutes apart.
Without the noise
One alert, not twenty.
- A quiet period after each alert. After a rule fires it stays quiet for 30 minutes. Several matching runs in that time become one alert for the most recent.
- Budgets once a month. The budget alert fires at most once per calendar month.
- Tests stay out of it. Test runs and evals never trigger a run or event alert, so checking a change does not page anyone.
What you get
Alerts that are set up, and stay tidy.
- A budget alert from day one. Every workspace comes with a Monthly budget exceeded rule, so setting a budget is enough to hear when you cross it.
- When each rule last fired. The rules list shows it, and a rule can be paused and resumed from the list.
- Unlimited rules. Every plan includes as many rules as you need. Owners and admins create and change them.
- Rules from code. List, create, update and delete rules with a workspace key through the API, or from Claude, Cursor or any MCP client.
Go deeper
How it works, in detail
More in Observe
See what every agent did, what it cost, and when it breaks.
Give one job to an agent this week.
Start with one repetitive workflow, gate the sensitive step behind your approval, and read the first run end to end.