rushour0//blog

Field notes · agent infra & a Game Boy game studio

Building an agent engine
by shipping a game on it.

Five essays on what it actually takes to make an LLM agent cheaper and more reliable — written from the commit log of fabri and ludexel, an engine and the product that keeps breaking it.

5 posts · read in order or jump in

01·build-in-public·6 min

I built the engine and the product at the same time — on purpose

Two repos at once — an agent engine and a Game Boy Advance game studio built on it. The product is the only honest test of the engine.

read →
02·agent-infra·6 min

Frozen prompts are the bug

A system prompt shouldn’t be hand-written and frozen. It should grow from the traces of what the agent actually did.

read →
03·cost / COGS·7 min

Orchestration is a cost decision, not a topology diagram

The interesting multi-agent question isn’t how agents talk — it’s the cheapest role for each step, and how you stop the bill from leaking.

read →
04·retrieval / IR·6 minstart here →

You can't improve retrieval you never measured

I built a self-improving memory system for weeks on vibes, then built the ruler. The first measurement showed two shipped things doing nothing.

read →
05·reflection·6 min

What building a second project taught me about the first

The fastest way I found to improve an AI agent wasn’t tweaking the agent — it was building something real that needed it.

read →