Kubernetes AI SRE Agent Finds an Unnoticed Crash Loop

You ask for a routine health check and expect a clean baseline. What came back was a pod that had restarted 788 times, unrelated to the question.

AURA is connected to a Kubernetes cluster and to Prometheus through read-only MCP servers, running as one coordinator with two specialized workers. The prompt is one sentence: check the health of the cluster, and confirm whether all the pods are running.

What comes back is not a baseline. AURA names the state as CrashLoopBackOff and attaches the restart count to it.

Partway through, some of the metrics work hits a provider rate limit. Rather than filling the gap, AURA states that the analysis is incomplete. The whole run is inspectable end to end: the plan, the workers assigned to it, and every tool call made.

A second question, about a suspected memory issue, runs two workers and 16 tool calls and returns nothing. The pods are running, ready, and at zero restarts, and the metrics do not support the theory, so AURA says so rather than producing a cause to match the question.

AURA reads and reports here. Both MCP servers are read-only and the run ends with findings and evidence rather than an applied change.

Chapters

0:00 The question, and the setup

0:18 A routine health check turns up a crash loop

0:35 Reporting an incomplete analysis instead of guessing

0:50 A second question, and a finding that is not there

1:09 From an unexamined cluster to a specific issue

#AISREAgent #Kubernetes #Prometheus

Transcript

The question, and the setup

0:00 What does an SRE agent actually find in a broken Kubernetes cluster?

We connected AURA to Kubernetes and Prometheus through read-only MCP servers. One coordinator, two specialized workers.

A routine health check turns up a crash loop

0:18 Then we asked something deliberately boring: "Check the health of my cluster. Are all pods running?"

We expected a baseline. Instead, a pod in CrashLoopBackOff. 788 restarts. We never asked about it.

Reporting an incomplete analysis instead of guessing

0:35 Part of the metrics work hit a provider rate limit. Instead of guessing, AURA said the analysis was incomplete. That matters.

The whole investigation is observable. The plan, the workers, every tool call.

A second question, and a finding that is not there

0:50 We pushed further with a suspected memory issue. One question triggered two workers and 16 tool calls.

This time, AURA didn't find the problem because it wasn't there. Running, ready, zero restarts. The metrics couldn't prove otherwise, so it said so.

From an unexamined cluster to a specific issue

1:09 One open question took us from an unexamined cluster to a specific issue, evidence attached.

This is AURA from Mezmo. To learn more, you can read the full blog article on what an AI SRE agent actually finds when you point it at a broken Kubernetes cluster at mezmo.com.

AURA use cases

Table of contents