Usually, if the issue is familiar to the LLM, if it knows of it and there is sufficient context for it in the LLM’s training data, it’s insanely good at reaching the correct hypotheses quickly. That is, when the problem space overlaps with the LLM’s knowledge base. For example, a pod CrashLoopBackOff due to the wrong image pull or memory limit can be easily spotted by an LLM. However, when the issue is novel, the chances of an LLM getting RCA right today are close to zero.
I do like the metaphor of MCP as a USB bus, but I think that far oversimplifies the matter presuming it is reliable consumption of that data. End of the day I find it far more useful to provide the exact context I feel is needed to solve the problem. Else I get made up nonsense.
PS This quote is gold. AI can only solve what it has SEEN. So good luck on your novel production crashes.
Quote Citation: SigNoz, “I built an MCP Server for Observability. This is my Unhyped Take | SigNoz”, 2025-07-18, https://signoz.io/blog/unhyped-take-on-mcp-servers/