Before you start
- Turning the health check on or off for the organisation needs a role that manages the organisation, such as owner or admin. Anyone else sees the switch but cannot change it.
- Each cluster has its own switch, which is on by default. A cluster is checked only when both switches are on.
Turn on the health check
1
Open Autonomy and approvals
Go to Ankra AI → Autonomy and approvals (
/next/organisation/ai-autonomy) and scroll to Finding problems on its own.2
Switch it on
Turn on Check cluster health on its own. It is off by default.The switch now reads on, and the settings below it become active.
3
Tune the check (optional)
- Restarts before a pod counts as unstable - leave it empty to use the platform’s default.
- Hours that restarts are counted over - only restarts within this many hours count. Leave it empty for the default.
- Leave out system namespaces - skips
kube-systemand the other system namespaces when checking workloads.
Leave one cluster out
- Go to Infrastructure → Clusters → <your cluster> → Settings → Ankra AI (
/next/clusters/<cluster-id>/settings/ai). - In the Ankra AI insights card, press Turn Ankra AI insights off, then confirm with Turn Ankra AI insights off in the dialog.
Where the problems it finds go
The health check does not have a page of its own in the portal. What it finds reaches people through the places they already work:
With the health check off, nothing is filed or sent from it, and no ticket or fix is started from a health check. You can still ask Ankra AI about a cluster at any time with
⌘J.
Alert-triggered analysis
When an alert fires, Ankra AI writes an analysis from what Ankra has recorded about the alert: the rule that fired, the resource it names and, where there is one, the failed job or operation behind it. The analysis does not read live state from the cluster, so for current pod status, events or logs, ask Ankra AI about the resource. The analysis is attached to the alert’s notification. To read it, open the notification from the bell (/next/notifications/<id>). Its AI analysis section holds:
- A summary of what happened.
- Likely cause - the underlying problem behind the symptom.
- Recommended actions - what to do next.
- Read full analysis - the complete write-up.
Which model reads it
Neither the health check nor the analysis on an alert follows the provider your organisation chose under Ankra AI → Models and provider.- The analysis on an alert always runs on Ankra’s own access to Groq. It is sent the alert and rule names, the names of the resource and the cluster, the failed job’s error messages and the start of its command output, and the Kubernetes events recorded with a failed health check. It is not sent pod logs.
- The health check runs on Ankra’s own access too: Groq first, and a Claude model if that fails. It counts against your plan’s monthly AI allowance. Once the allowance is used up, it runs on your organisation’s own key if you added one.