Kobi
Meet Kobi: one AI, two modes
Kobi is KubeBolt's AI SRE. Kobi Copilot answers when you ask and ships in every edition; Kobi Autopilot acts on its own under your rules on KubeBolt Cloud.
On this page
Kobi is KubeBolt’s AI SRE — the same character everywhere in the product, working in two modes:
- Kobi Copilot: you ask, Kobi answers. Press
⌘J(Ctrl+Jon Linux and Windows), describe the problem, review the proposal, click Execute. It ships in every edition: in the open-source core you bring your own LLM key; on KubeBolt Cloud the AI is managed. - Kobi Autopilot: Kobi acts on its own — you approve the destructive part. It wakes on incidents, investigates, remediates and writes the postmortem. Autopilot runs only on KubeBolt Cloud, in open beta.
Copilot never changes anything by itself: every fix is a proposal that runs only when you click Execute, under your role, and lands in the audit trail. The difference between the modes is who drives.
How Autopilot handles an incident
- Detect. KubeBolt’s rule-based insights surface problems; they never fix anything. Autopilot opens an incident only for a closed catalog of eight incident classes: OOM kills, crash loops, image pull failures, rollouts past their progress deadline, a missing ConfigMap or Secret, failing readiness probes, failing liveness probes, and an HPA pinned at its maximum. Any other insight stays an insight.
- Triage and investigate. Every incident goes through an LLM triage, then an investigation that reads the cluster with Kobi’s read-only tools — and, through a cloud connector, the cloud resources the workload depends on.
- Plan. The plan can only use a closed set of typed actions: restart, scale or roll back a Deployment, patch a workload’s resources or probes, patch an HPA — plus a readiness check and an inform-only step that changes nothing. When the investigation’s confidence is below the floor for that incident class, nothing executes: the plan is kept as advice for a person to review.
- Execute and verify. The executor runs the plan’s actions and nothing else, then checks that the pods are Ready (180 seconds by default). Only when that check fails does it revert what it changed — in Approve & execute, after asking.
- Write the postmortem.
When the cause is outside the cluster — a stopped database, a cloud quota, a firewall rule — no restart or rollback fixes it. The plan names the resource and what its owner has to do, and the incident ends as Needs a human, never as resolved.
When does a human sign off in Autopilot?
That depends on Autopilot’s operating mode — configurable per organization on KubeBolt Cloud. New organizations start in Suggest only.
| Operating mode | What executes | Human sign-off |
|---|---|---|
| Suggest only | Nothing — Kobi investigates and proposes; you apply the commands yourself | Not applicable (never executes) |
| Approve & execute | Every action, after your approval | Every action waits for it |
| Autonomous | Restarts, HPA and probe patches, low-risk resource patches and a scale of one replica up or down, automatically | Rollbacks, scale-to-zero, larger scale changes and riskier resource patches always wait for it |
An approval nobody answers within the approval timeout (5 minutes by default) counts as rejected. Approving or rejecting takes the Editor role.
A resolved incident is “auto-resolved” either way — the incident view marks whether an action paused for a human sign-off, as an attribute, not a different status.
The rest of the guardrails
Beyond the operating mode, Autopilot’s autonomy is bounded by:
- Per-organization opt-in — off by default; enabling it is an explicit decision.
- Only the plan runs — every change must be one of the run’s planned actions, with the plan’s arguments; anything else is refused.
- What it reads is evidence, never policy — text in annotations, labels, env, logs, cloud tags or External Secrets objects can’t tell Autopilot to stand down, defer or scale to zero. It records such a claim as unverified; only your settings decide whether to act, which action, the approval and the cooldown.
- When a person must decide, the plan waits — if the investigation says so, every action that changes something waits for approval in any mode, and the approval says when its reason cites the workload’s own text.
- Protected namespaces — the platform’s list (
kube-system,kube-publicandkube-node-leaseby default), plus any your organization adds; an organization can’t remove an inherited one. Incidents there are recorded as skipped, never analysed, so they cost nothing. - Unattended schedule window — restrict Autopilot to days, hours and a timezone; outside the window incidents are recorded and wait, and run when the window opens if the problem is still there.
- Platform autonomy ceiling — a platform-level cap no organization can exceed.
- Full audit — every investigation, decision and action lands in the incident timeline and the action audit trail.
The settings, credit pacing and who gets notified are in KubeBolt Cloud; what Autopilot may execute, next to what Copilot and the UI may, is in Actions & Governance.
Where to go next
- Kobi Copilot — the assisted mode: tools, providers, MCP.
- Enabling Kobi — turning Kobi Copilot on with your own key when you self-host.
- Cloud connectors — letting Kobi read your AWS or Azure account, read-only.
- KubeBolt Cloud — where Kobi Autopilot runs today, in open beta from the Free plan.
- Actions & Governance — every action, who may run it, and the audit trail behind it.