Setup
This is the kit behind the work: what it runs on, which models do what, and how the agents are held together. I update it when something changes.
Hardware
MacBook Pro, M2 Max
My main machine. I do nearly all of my work from a terminal on it.
Mac mini
Does the work that needs a screen: running apps, emulators and browsers so the agents don’t take over my laptop.
A VPS
Always on. The agents that watch for email and messages run here, so they keep working when the laptop is shut.
Meta Quest 3
For VR work. The agents install, test and drive apps on it.
Even Realities G2 glasses
Smart glasses I write apps for and use to talk to an agent.
Android phone
What the agents use to test mobile apps and do the things that only work from a phone.
Models
One model does the thinking. The others are for things it can’t do well itself.
| Model | Used for | As of |
|---|---|---|
| Claude Opus 5.5 | All agent work: coding, writing, operations. | Nov 2026 |
| Gemini | Watching video, frames and audio together. | Nov 2026 |
| Nano Banana (Gemini) and fal | Making and editing images. | Nov 2026 |
| An open-source depth model | Turning 2D video into 3D for the Quest. Runs on the laptop. | Nov 2026 |
The harness
The agents run in Claude Code. I’ve taught them hundreds of skills: short playbooks, with scripts, for one kind of job each.
Hooks are the guardrails. A hook is a small program that runs before or after an agent does something and can stop it, for example from pushing code that fails its tests.
Daemon is the system that wakes an agent when something happens, such as an email or a message arriving. That is how support gets answered while I’m asleep.
The agents use command-line tools instead of MCP servers. A command is cheaper for a model to read and easier to test, and I can run the same command myself.
How much this adds up to each week is on the stats page.