flagship · Journal of Grid Computing 2026
A human-governed AI SRE for Kubernetes.
You ask a question in plain English; it investigates your cluster with real tools —
kubectl, Prometheus, Loki — explains what it found, and, with your approval,
fixes it. It acts on the cluster, not just talks about it — and every mutating
action pauses for explicit human sign-off before anything runs.
// the problem
Running Kubernetes at scale is a continuous operational burden. Failures are diverse — crash loops, OOM kills, image-pull errors, pending pods, missing endpoints, admission rejections — and the knowledge to resolve them is spread across the cluster API, metrics, and logs. On call at 2 a.m., an engineer bridges those tools by hand, correlates in their head under pressure, and hopes the fix works. Nothing is remembered, so the same fault pages someone else next week.
First-generation LLM ops assistants share three structural weaknesses:
Reactive. It does nothing until a human notices a problem and types a question — a slow-burn failure at 3 a.m. is invisible until someone looks.
Uniformly expensive. It calls one large model on every turn, even for trivial "is this even a problem?" triage a compiled rule could answer for free.
Unauditable. When an agent touches a cluster, a chat transcript is not an audit log — you can't prove the record wasn't altered after the fact.
Continuous perception. Zero-token detectors watch the event stream and recognise known failures at no LLM cost — catching problems before they page anyone.
Cost-aware reasoning. A tiered "Cortex" reserves the most capable model for final synthesis and uses lighter compiled steps for triage.
Provably auditable. A hash-chained, tamper-evident flight recorder makes every decision deterministically replayable, with rollback points on any change.
// what it is
KubeIntellect inverts conventional Kubernetes tooling: instead of you knowing what to look for
and the tool fetching it, you describe what's wrong and it figures out where to look —
correlating logs, metrics, events, and configuration across many signals in parallel. Operators
don't need to know kubectl syntax, PromQL, or LogQL.
For a simple question, you get a direct plain-English answer. For a genuine outage it runs a root-cause analysis: it fans out specialist agents — one each for the pod, its events, its logs, and its metrics — and synthesises their findings into one structured root cause, with supporting and conflicting evidence, and a recommended change shown as a dry-run diff you approve before it applies.
// architecture
A planner/router decides, per query, whether to answer directly, run a targeted single-resource deep dive, or fan out a full root-cause investigation across concurrent domain specialists.
Destructive operations pause mid-execution for explicit approve/deny. Cascading-blast actions (delete namespace/PV/CRD, drain, resource changes) always require confirmation — even in auto-approve mode.
readonly → operator → admin → superadmin, enforced at the tool layer with namespace guardrails. It won't read Secrets or ServiceAccounts, so the agent cannot exfiltrate credentials.
A tamper-evident, deterministically replayable audit log; mutations arm rollback points and produce grounded, source-cited incident postmortems.
In the architecture described in the paper, KubeIntellect generated new Python tools at runtime, validated them with AST safety checks and tested them, and could graduate a proven tool into the codebase via an automated pull request. Those generated tools executed in-process behind AST validation and were not sandboxed. The capability was evaluated in the paper and dropped when the system was simplified for production; the shipping implementation works from a fixed, audited tool set.
Compiled, zero-LLM-cost detectors watch streaming events; a graduated autonomy ladder lets it open investigations unprompted, strictly within human-set per-namespace bounds and allowlists.
// evaluation
Validated on a live cluster against a controlled, reproducible fault-injection corpus — each scenario ships an injection manifest, a gold expected diagnosis, and a recovery-verification script — scored by an independent judge model and fully regenerable from the released harness. KubeIntellect is open source and pre-commercial; every number is reproducible.
// white paper
A decision paper for platform, SRE, DevOps, and cloud-native decision-makers — continuous perception, tiered reasoning, and auditable autonomy for LLM-driven Kubernetes operations.
Download ↓// stack & deployment
Multi-provider LLM support and ready-made overlays for Kind, EKS, GKE, AKS, and ACK. Packaged on
PyPI as kubeintellect and kube-q; deployable via a guided init wizard,
Docker Compose, or a Helm chart to any cluster on any cloud.