Field notes · AI agents & a Game Boy Advance game studio
Five essays on AI agent orchestration, cost, and retrieval — what it actually takes to make an LLM agent cheaper and more reliable while shipping a 2D Game Boy Advance game studio on top of it. Written from the commit log offabri andludexel, an agent engine and the product that keeps breaking it.
5 posts · read in order or jump in
Series
Five essays on AI agent orchestration, cost, and retrieval — building an AI agent framework (fabri) and shipping a 2D Game Boy Advance game studio (ludexel) on top of it.
Two repos at once: an AI agent framework and the Game Boy Advance game studio built on it. Notes on building an agent engine and shipping a 2D game in public — the product is the only honest test of the engine.
read →A system prompt for an AI agent shouldn’t be hand-written and frozen — it should grow from the traces of what the agent actually did. Notes on prompt engineering and self-improving agent memory.
read →The interesting multi-agent orchestration question isn’t how agents talk to each other — it’s the cheapest model for each step, and how you stop an AI agent’s LLM cost from leaking.
read →I built a self-improving RAG memory system for weeks on vibes, then finally built the eval. The first recall@k measurement showed two shipped retrieval features doing nothing.
read →The fastest way I found to improve an AI agent wasn’t tweaking its prompt or code — it was building a real product, a 2D Game Boy Advance game studio, that actually needed it.
read →