I’ve been meaning to write about this stuff for a while.

I run a small fleet of AI agents. Five of them right now — named Bob, Bill, Riker, Bluebells, and Mario. They live in containers on a server in my house, connected to Discord, and they do things: research, media management, code review, general conversation. Some of them have personas, all of them have tools, and none of them are plug-and-play. Getting agents to actually work reliably takes real infrastructure.

The term for the person who does that work seems to be settling on operator. Anthropic uses it officially in their usage policies — the principal hierarchy goes Anthropic → Operators → Users. An operator is someone who accesses model capabilities through an API and deploys them in a specific context. OpenAI actually named one of their products “Operator.” It’s becoming vocabulary.

I think “wrangler” is also a good word for it. Maybe more honest about what the day-to-day feels like.

What does operating agents actually involve? From my experience:

  • Configuration that never quite ends. Which model handles what, which channels the bots respond in, what tools they have access to, what the fallback chain is when the primary model is down.
  • Infrastructure work. Networking, secrets management, container orchestration, health monitoring. My setup is NixOS with sops-nix and Podman — fully declarative, which helps, but it’s still a lot of surface area.
  • Prompting as architecture. System prompts aren’t just instructions, they’re load-bearing. Getting an agent to behave consistently requires thinking carefully about what it knows about itself and its context.
  • Debugging things that shouldn’t be possible. Stale iptables DNAT rules. Models that work perfectly in isolation and fail when called as subagents. Bots that go offline because a container restarted and left orphaned network state.

This blog is where I’m going to write about that stuff. Infrastructure decisions, things I learned the hard way, thoughts on where the operator role is heading.

The domain is parks.place — Parks is my last name, and “place” felt right. My demo subdomain is demo.parks.place, which I plan to use for showing work things at work without having to explain what warren-desktop.tailscale means.

Welcome. More soon.