- A LangGraph workflow: Claude proposes the evolution and a reviewer validates it. If something is wrong, it sends the errors back and asks for a fix. In one real run, Claude tried to set armor to 5 when the cap was 0.5, and the reviewer caught it.
- Structured outputs with LangChain and Pydantic-validated contracts. The AI never touches the game directly.
- Hard limits and a plan B: with no network or a failing model, my own rules decide, and the game keeps running.
- No frozen frames: the AI thinks on a background thread during the break between waves.
- Observability with LangSmith: every decision is traced.
- Automated tests that cover every AI path without spending tokens.
- Bilingual: Spanish and English, including the explanations the AI writes.