I’m building NoPager, an open-source incident-triggered AI operations engineer.

The original idea was simple:

Production breaks. NoPager wakes up — not you.

But one of the first external reviewers pointed out a serious problem in the early architecture:

if the AI worker has access to docker.sock, then all the “bounded action” rules above it are ultimately application-level promises.

So I changed the architecture.

The ordinary NoPager worker now has:

no Docker socket no Docker group membership no Docker CLI no free-form shell execution

A small host-side Rust helper owns the privileged boundary instead.

The AI can reason over incident evidence, but the execution layer only accepts a narrow typed request against one enrolled target.

The current operations loop is:

health failure → incident → AI triage → Safe Mode approval → one bounded restart → independent verification → resolve or escalate

The part I’m becoming most interested in is this:

AI should be allowed to think broadly, but act narrowly.

That seems especially important once AI agents are allowed anywhere near production infrastructure.

NoPager is also event-driven rather than continuously running an expensive model. Lightweight monitoring stays on 24/7; the AI wakes only when an incident actually needs reasoning.

GitHub: https://github.com/nopager/nopager

I’m currently looking for people willing to try it on a disposable or non-critical Docker service.

I’d especially like feedback on one question:

What production action would you trust an AI agent to perform automatically — and what action would you always keep behind human approval?