THE NOTEBOOK
Field notes from building real AI systems.
Architecture, harness design, and what happens when models meet production.
Quiet notes on working systems.
05 NOTES NEWEST FIRST
Loop Engineering replaces you — the person prompting the agent — with a system you design. The four layers that led here, what a loop actually is, the principles that keep one honest, and a worked implementation you can run today.
The evaluator — the mechanism that grades how well an agent performs — is the most postponed part of the harness, yet the part that decides whether you reach production and can ride the coming self-improvement wave. Why it belongs at the start of design, not after deployment.
Putting a rule once in CLAUDE.md vs. re-pasting it every turn — measured against the real claude --print. The fixed cost of framing, the re-send penalty multiplier, and how conflict resolution splits by model (Sonnet stays faithful, Opus leans to recency).
The meta-harness skeleton survives the move from an individual user to an enterprise — but several design assumptions flip. A two-layer scope design for self-improvement loops you can actually operate.
What happens when you build a personal AI that can observe and modify its own harness — and how to implement it without letting it spiral out of control.
The ink dries, and the next experiment begins.