Why your SRE team is drowning in alerts, even with AI

The CNCF just published a sharp explainer on a problem every Brown and Black-led ops team knows by heart: you can’t automate what you don’t understand.

It’s called the “4-body problem.”

The four bodies are:

1. The alert that wakes you up at 2 a.m.
2. The context of what actually happened
3. The tools your team already owns
4. The human decision of who to page

Most SRE teams treat this like a pure AI problem. They plug in an LLM, feed it a runbook, and expect it to fix the outage. It doesn’t. It hallucinates. It pages the wrong primo. It escalates a false positive to the VP.

The CNCF’s take is simple and brutal: autonomous operations fail when the AI has no context. No shared memory. No understanding of your specific architecture. No awareness of who’s on call and why they’re already stretched thin.

The article walks through real failure modes — not theoretical ones. An AI agent misclassifies a network blip as a database crash. It tries to “fix” it by rolling back a deployment that was never broken. It creates a new outage while trying to solve the old one.

The fix isn’t a better model. It’s better context layers:

- Shared state between alerts and runbooks
- Human-in-the-loop checkpoints for high-risk actions
- Clear ownership so the AI knows who to defer to

It’s not a product launch. It’s a framework. And it’s free.

Why this matters for us: Every team running side-hustle infrastructure — the cousin’s e-commerce site, the auntie’s community clinic booking system, the neighborhood food delivery app — is already living this 4-body problem. When the AI agent makes a mistake, it’s not a “system error.” It’s someone’s rent money on the line. We need ops that respect our context, not pretend it doesn’t exist.