DOP 360: What Is an AI SRE?

#360: Assume your production systems go sideways at 3:00am. Your phone doesn’t ring. You wake up to a clean incident report - the system caught the problem, fixed itself, and left a Slack note saying it handled things at 3:15. Would you trust it? Viktor wouldn’t, not at first. Same as a new hire, a new tool, a new anything. Trust gets earned. A few months in, if it hasn’t messed up, then maybe.

So what’s actually new? We had self-healing in decades past, and Kubernetes brings a crashed pod back without anyone awake to watch. Birol Yildiz, co-founder and CEO of ilert, puts the line where it belongs: the old stuff runs on predefined patterns - known problem, known fix, automate it. The new part is diagnosing a root cause nobody wrote a rule for and then knowing what to do about it. And no, ilert isn’t training models on your data. The agent runs on foundational models from the frontier labs, feeds them the same context any SRE would pull - telemetry, recent changes, the health of your infra - runs the investigation loop, and hands you a root cause with the evidence attached.

Then the real argument starts, and it’s about knowledge. Viktor keeps circling the blind spot he sees at every company: the wiki, the runbooks, the tribal stuff nobody keeps current. Birol’s answer is blunter than you’d expect from a guy selling this - dump your Confluence and your hundred stale runbooks into the agent and you cause more damage than results. Code is the documentation that’s actually true. ilert’s agent doesn’t even read runbooks. It builds a live service topology straight from tracing data and rebuilds it when things drift.

What happens when a bad incident month turns your $100 bill into $10,000 - Birol’s answer is a credit budget with a human paged when you blow past it. Who’s accountable when the agent makes it worse? Has to be a human. Agents are tools. You don’t get to blame your AI coding agent for shipping garbage code, and ilert doesn’t get to blame the model either. Trust is binary - five good runs and one bad one, and you strip its write access. But even in observe-only, the value is real, because the time sink in an incident was never the fix, it was finding the needle. Cut root cause from 45 minutes to a few. And the closer nobody wants to hear: if you skipped observability and fifteen years of homework, an AI SRE isn’t going to save you. You can’t skip the parts you skipped.