AI SRE Agent Audits Runbook Coverage and Opens the PRs: AURA
3 a.m., the pager fires, and the runbook describes a service that shipped three versions ago. Ask the agent what the cluster actually has instead.
Learn more about AURA → https://www.mezmo.com/aura
Get started today → https://github.com/mezmo/aura
Get deployment help → https://www.mezmo.com/contact
Runbooks go stale because clusters change faster than documentation does. Every deploy, every new service, every renamed alert widens the gap between what is running and what is written down.
Asked to investigate the apps in the cluster against the runbook repo, AURA works both sides at once, enumerating what is live and reading what is documented. In this cluster the answer was 38 workloads across 16 namespaces and zero runbooks, just a README and an empty template. The three services failing at that moment, all OOMKilled, had nothing. AURA rechecks live state, authors three complete runbooks against the repository's own template, and opens a pull request. 700 lines, grounded in what it observed.
That pull request surfaces a second gap. Only three Grafana alert rules exist and none fire on OOM kills, memory against limit, or restart rate, so every one of these outages was caught by a person rather than an alert. Asked how to resolve it, AURA returns three failure modes and three PromQL rules for the alerting module Terraform already provisions. No new repository, no new exporter.
AURA proposes and does not merge. Every mutating call stops for approval, the branch, the file, and the pull request itself, and a person reviews both PRs like any other change. Re-running the same investigation afterward reports three merged runbooks and five live alert rules, one of them already tracking the frontend's memory climb.
The setup is two MCP servers, Kubernetes and GitHub, and two workers. Cluster inspector holds read-only K8s tools; GitHub analyst holds the repository tools plus the three writes it needs to branch, commit, and open a pull request. The gate is one list: create, update, push, merge. Anything that changes state.
For more detail, visit
https://www.mezmo.com/videos/ai-sre-agent-audits-kubernetes-runbook-coverage-aura
0:00 The runbook is three versions out of date
0:12 Why runbooks drift faster than clusters
0:27 One prompt against both sides of the gap
0:55 The finding: 38 workloads, zero runbooks
1:16 AURA authors three runbooks and opens a PR
1:41 The alerting gap the pull request surfaced
2:18 The second PR, gated on every mutating call
3:00 Re-running the audit
3:21 The other direction: incidents teaching runbooks
3:40 The setup: two MCP servers, two workers
#AURA #Runbooks #AISRE