It's 3am. Something is broken. Who do you trust to find it first, yourself or the AI?

In this session we break production live, on purpose, and race three ways of fixing it: an engineer digging through logs the way we always have, the AI assistant your cloud console already ships, and an open source SRE agent wired over the Cloud Logging and Monitoring APIs, the kind you can clone and run on Monday. Same incident, three paths, clock running.

We'll break three things. A bad deploy throwing 5xx, where the AI wins in seconds and you understand immediately why this matters. A misconfigured IAM permission cascading into 403s downstream, where correlating logs and configuration in parallel genuinely beats human tab switching. And then a trap: a symptom that points confidently in the wrong direction, where the model produces a beautiful, plausible and completely wrong hypothesis.

That last one is the whole point of the talk! These tools hand you ranked observations and hypotheses, not verdicts, and the well designed ones never touch your environment on their own. Learning to read that output as evidence instead of as an answer is the actual skill, and nobody teaches it.

Everything runs on open source and on a throwaway project, so nothing you see depends on a licence you might not have. You'll leave knowing where AI genuinely compresses your time to resolution, where it will confidently mislead you, and how to wire an agent into your own observability stack without handing it the keys to production.

It's 3am and something is broken. Who finds it first, you or the AI? We break production live and race three responders: a human in the logs, the assistant built into your cloud console, and an open source SRE agent you can clone and run yourself. Three failures, and one of them is a trap where the AI is confidently, beautifully wrong. Come and learn where AI really cuts your time to resolution, and where it lies to you with a straight face!