The Thesis

The infrastructure around the model matters more than the model itself. Models get better every quarter. Prompting tricks get obsoleted with each release. But the harness you build around them compounds. Each piece -- the sandbox, the review loop, the skill system, the session history, the coordination layer -- makes every future session better, not just the current one, and not just for you but for every agent in the system.

The argument has four parts. The boring stuff is where the actual leverage is: conventions, file structure, handoffs, verification. Not clever prompts. The loop explains why this compounds: sessions leave artifacts, artifacts feed future sessions, and each round starts with more context than the last. You build this infrastructure by starting manual and automating later, because premature automation encodes the wrong patterns. And the compounding only works in agentic mode, where the agent operates inside a persistent environment that keeps artifacts around. Autocomplete gives you a faster session. The harness gives you a faster next session.

The Boring Stuff — the split between what the agent does and what I do.

The Loop — why compounding knowledge is the real bet.

Start Manual, Automate Later — the recurring pattern.

Assistive vs Agentic — why the loop only works in agentic mode, and why the skeptics are half-right.