What ARGUS is¶
ARGUS is a self-hosted platform that detects Kubernetes faults, investigates them with an AI model, and proposes a reviewed fix. It is built for the gap between "an alert fired" and "the cause is understood".
The problem it addresses¶
Kubernetes tells you a great deal about symptoms and almost nothing about causes. A pod is
in CrashLoopBackOff; the cause is a missing ConfigMap key, a memory limit set too low, a
volume that cannot attach to two nodes at once, or an image tag that does not exist. The
symptom is on screen in seconds. The cause takes a person with cluster access, time, and
familiarity with that particular workload.
ARGUS does the middle part. It gathers the evidence a person would gather, states what it concludes and how confident it is, shows the evidence behind the conclusion, and — where a fix is expressible — proposes one.
What makes it different from monitoring¶
Monitoring tells you that something is wrong. ARGUS is built around the three steps after that.
It groups symptoms into problems. A failing deployment produces dozens of events across several pods. ARGUS collapses those into one incident per affected object, and links incidents that share an owner or reference each other's resources.
It investigates with evidence, not guesses. The investigation runs against your live cluster through a bounded set of read-only tools — describing objects, reading logs, listing events, resolving owner chains, asking whether an object is managed by a GitOps controller. Every step it took is recorded as a causal chain you can open.
It knows where a fix belongs. If a broken object is managed by Argo CD, Flux or a Helm release, changing the cluster directly is a stop-gap that the next sync reverts. ARGUS works that out from the cluster's own ownership facts before it proposes anything, and routes the fix to your repository instead.
The approval boundary¶
Every change ARGUS can make to a cluster is an approval away. The set of changes it can make at all is a fixed list of typed actions — scale a workload, restart pods, patch resource limits, revert a container image, delete a named object, create a limited object, open a pull request. It has no shell, and no way to run a command you did not anticipate.
A new installation executes nothing. Agents install read-only, and remediation is enabled per cluster and per namespace by whoever installs the agent.
The honest position on confidence¶
An investigation reports a confidence figure. It is a signal from the model, not a measurement, and ARGUS treats it that way: the confidence is displayed next to the evidence, never instead of it. Where the investigation reaches a bound without concluding, it says so and reports what it found rather than inventing a conclusion. Where a cause is clear but no supported action expresses the fix, it says that too — that outcome is common, and it is a result rather than a failure.