House doctrine for designing LLM agent systems: safety enforced by the harness and tools rather than model goodwill, tools that execute and return results, minimal loop code, interface-first component seams, typed config, deterministic evals over LLM judges, and POC scope discipline. Use when designing or reviewing any agent, tool-use loop, eval harness, or agent-backed CLI, in any framework. Do NOT use for Pydantic AI framework mechanics (use building-pydantic-ai-agents), generic design principles (use review-design), or Go agent-loop implementation details (use agent).
