Five questions to answer before designing an AI agent architecture

The author argues that AI system architecture should begin with analyzing the problem and the available data, not with writing code. Five questions guide the design: who performs each task (an AI agent, a Python program, or both), where the work runs (a remote VPS, a local computer, or elsewhere), what data sources and knowledge base are available, what outcome is desired, and how that outcome is delivered. The author illustrates this with an AI sales assistant for Instagram Direct, where Python enforces a conversation script while Claude handles natural replies drawn from a knowledge base. Collected information is sent to a doctor as a Telegram card, and a human confirms the appointment before it is created. The key point is that these responsibility and data decisions are made before any implementation begins.