Noisy Neighbors, Not Bad Tenants: Can your AI SRE tool Tell the Difference?

A story about dogfooding an AI SRE tool (aiHelpDesk) against a Kubernetes noisy-neighbor memory pressure scenario, which uncovered three platform bugs rather than just improving playbook content. The bugs all shared a common root cause: letting LLM self-reports substitute for values the system already knows authoritatively. Key lessons include removing ambiguous agent-selection prompts, threading namespace values server-side instead of relying on model output, and isolating test harness flags from diagnosis prompts. The post argues that full audit trails and transparent, versioned YAML playbooks are what make AI SRE tooling trustworthy and debuggable, contrasting this with black-box vendor approaches.