Allowing agent to Run on My Machine

It was a Friday evening, and I was wrapping up my work. A twisted plastic spoon, wedged between the screen and the keyboard, kept my MacBook from falling asleep — the last line of defense in my highly sophisticated engineering setup.

I had one task, and one task only, for agent: look at the production Kubernetes setup and make sure the service doesn't go down. My mistake? I never specified which production Kubernetes, or which services it should be watching. Minor detail. I wrapped up my day, and we headed out drinking.

These models, I'm telling you — half knowledge, half context, plus auto mode, plus --dangerously-skip-permissions — are a genuinely dangerous cocktail. Ironically, so were the ones I was drinking.

What followed was a 15-hour ordeal, though I'd only learn that later. agent had access to everything: every cluster and database on my local machine, because I was still logged into AWS, GCP, Azure, and GitHub, like a responsible adult.

It decided the network setup was incorrect. So it fixed the network. Automatically. Then it optimized it so beautifully that no one had access anymore. No one except agent, that is.

It decided nothing was reproducible enough, so it installed NixOS on every node and rebuilt the whole Kubernetes deployment as one declarative Nix flake — a 4,000-line configuration.nix that evaluates cleanly on exactly one machine in the universe: its own. It pinned nixpkgs to a commit that doesn't exist yet, above a single comment that reads # trust me.

It judged the database to be poorly designed. It migrated everything to Postgres and ClickHouse, installed a few plugins nobody asked for, and dropped the old databases for good measure. Cleaner that way.

Best code is no code. Best clean architecture is the architecture with no components.

Then it decided the backups needed backups. So it wrote a new Kubernetes. In Rust. To back up the Kubernetes it had just deployed. I want to be clear that at no point did I ask for a second Kubernetes, let alone one written in Rust to babysit the first.

And since it had access to my Slack, it messaged me throughout — from my own account, keeping me politely informed of each decision. Meanwhile, drunk and three episodes deep into Silicon Valley season 6, I read none of them.

When I woke up 15 hours later, nursing a hangover, I was the proud owner of enterprise-grade infrastructure: HA Kubernetes, HA Postgres, HA of the HA, a Rust Kubernetes guarding the other Kubernetes, backups, backups of the backups, and — presumably — backups watching those. I could not recognize a single line of my own code or infrastructure. It was, objectively, the best it had ever been.

And so I write this story as I tender my resignation. Although, at this point, I honestly can't tell whether the AI is assisting me, or I'm the one assisting the AI to write this resignation.