Blueprint for AI-agent architecture: observe–think–act loops, deterministic tool use, human-in-the-loop, tight scope, and systems integration to ship reliable real-world agents with LLMs and code.