PARACODING: Everything Is Still a DNS Problem: Building Production AI Infrastructure by Supervising AI Agents

The room said DNS couldn't do it. He'd already heard that once.

In 2015, Scott McDonald built DNS failover for websites when the industry swore caching made it impossible. It worked — and got him acquihired into one of the biggest clouds in the world. Ten years later he asked the obvious question: does the old trick work for dying GPUs?

The answer was half yes — and finding the boundary of that half is the first story in this book. The second is bigger: McDonald barely touched the keyboard. He built a production AI-inference platform by directing AI agents — writing the briefs, granting the approvals, holding the credentials and the judgment — then used the very same method to write and publish the book you're reading.

Everything Is Still a DNS Problem is two case studies and one repeatable method for building real infrastructure by supervising AI agents instead of typing every line yourself. It's about where the boundary of that autonomy sits — which decisions the machine never gets to make — and what it feels like to be the human still holding the pager.

This isn't an AI hype book. Every claim is labeled — measured, reasoned, or unverified — and when a number couldn't be traced back to a log, it didn't get printed. The failures are enumerated in a public, runnable repository, including the bug that bit the project's own AI architect twice while it was documenting that exact bug.

Inside:

A two-tier HA/DR architecture for LLM inference that survives GPUs dying mid-request The first-token contract — how to fail safely in the middle of a streamed response The telemetry contract that makes silence mean death The operating model: briefs, approvals, one-way doors, and knowing when to grab the wheel Four appendices, drawn architecture diagrams, and a lessons list you can steal If you write software, run infrastructure, or just want to see what supervising AI agents actually looks like when the receipts have to add up — this one's for you.

Two case studies. One method. All the receipts.
And everything is still a DNS problem.