Know when scheduled work doesn’t happen
Investigate stale heartbeats, overdue schedules, stalled runs, and repeated failures in an agent supervision workspace.
When an agent fails quietly, nothing tells you
Recurring agents rarely have someone watching every run. Report the expected schedule and execution state so your team has a way to notice when the routine breaks.
- Mon08:12
- Tue08:19
- Wed08:26
- Thu08:33
- FriNo report
Separate health from delivery
A recent heartbeat means the agent checked in. It does not establish that the scheduled task started or the output was delivered. Investigate those conditions independently.
Acknowledge with context
Open the incident, inspect the related activity, and record that someone is investigating. Acknowledgement is a human response, not evidence that the underlying condition has recovered.
Build a practical review routine
Check open incidents at the start of the day. Prioritize the workflows other work depends on, inspect their last run, and verify recovery before relying on the next output.
The four conditions that open an incident
Incidents are evaluated from what your agents report, so they catch the failures that never produce an error message.
- No heartbeat for 10 minutes from an agent that is expected to be running.
- A scheduled run overdue by 15 minutes: the agent reported when it would next run, and that time passed without the run starting.
- A running job with no update for 30 minutes, neither to the run nor to any of its steps.
- Three consecutive completed attempts that failed.
When the condition clears (a heartbeat arrives, the overdue run starts, the run moves again, a run succeeds), the incident records the recovery. If the condition comes back, the same incident reopens and counts the occurrence instead of creating a duplicate.
Agents that only run on a schedule
A scheduled agent goes quiet between runs by design. If it has reported its schedule, Toolcaise shows it as idle until the next run is due, and a run that does not start becomes a missed-schedule incident rather than a false heartbeat alarm. An agent wrapped by Toolcaise Connect is idle while its machine keeps checking in. Either way, you hear about a nightly job that did not run, not about a nightly job doing what nightly jobs do.
Alerts where your team already works
Send incidents to the places people actually look. Each destination is verified before it receives anything.
- Slack
- Microsoft Teams
- Discord
- PagerDuty (US or EU)
- A webhook of your own, with signed deliveries
Choose the severity each destination receives. Email, chat and webhook destinations can also take a daily digest and a weekly report, so a quiet week still produces a record that someone looked.
Acknowledge, investigate, resolve
An incident moves through open, acknowledged, recovered and resolved, and every change is recorded with who made it. Acknowledging says someone is on it; it does not claim the problem is fixed. Resolving it while the condition is still active is allowed, but the incident keeps showing that the condition has not cleared, so nobody mistakes a closed ticket for a recovered agent.
Report a schedule so missed work is visible
Overdue-run detection needs to know when the next run is expected. Report the schedule with the agent’s telemetry: a name, a calendar (a cron expression or plain text such as “daily 06:00”), a timezone and the next expected run.
Frequently asked questions
How do I know if a scheduled AI agent didn’t run?
Report the schedule with the next expected run time. If that time passes by 15 minutes without the run starting, Toolcaise opens a missed-schedule incident and alerts the destinations you chose. A heartbeat alone cannot tell you this, because a healthy process can still skip its work.
Will I get alerts for agents that are idle between runs?
No. Agents that report a schedule are shown as idle until their next run is due, and agents Toolcaise Connect wraps are idle while their machine checks in. An incident opens when an expected run is missed, or when a Connect machine goes quiet.
Which alert channels are supported?
Email, Slack, Microsoft Teams, Discord, PagerDuty in the US or EU region, and signed webhooks. Each destination is verified before it is used, and you choose which severities it receives.
Does acknowledging an incident close it?
No. Acknowledging records that someone is investigating. The incident recovers on its own when the condition clears, and a person can resolve it. If the condition returns, the incident reopens and counts the new occurrence.
Related guides
How to spot an AI agent stuck in a loop, stop it with limits or a kill switch, end its whole process tree, and design runs so a stop is safe.
Read the guide How to detect silent failures in scheduled AI agentsWhy scheduled AI agents stop without an error, and how to catch it: expected-schedule checks, stale heartbeats, stalled runs, output checks and alert routing.
Read the guide A rollout checklist for putting your first AI agent into productionA step-by-step checklist for deploying an AI agent to production: define a good result, test failure paths, set limits and approvals, then decide go or no-go.
Read the guideConnect your first agent
Create a free workspace, connect the agent you already run, and watch its next run arrive. Free covers 3 agents, with no card required.