Building an AI agent from scratch in Python (www.leoniemonigatti.com)

A hands‑on tutorial walks through building a single-agent AI system from scratch in Python using direct LLM API calls (Anthropic’s Claude 4 Sonnet via the anthropic client) rather than an off‑the‑shelf agent framework. The author argues starting with raw API calls helps you learn the mechanics before adopting abstractions; the tutorial implements an Agent class with a system prompt (instructions), short‑term conversation memory, tool integration, and a run loop so the agent can hold multi‑turn context and call external functions. Practical setup details—storing ANTHROPIC_API_KEY in env/.env and the model string "claude-sonnet-4-20250514"—are included. Technically, the agent stores messages in a list of {"role","content"} pairs, passes tool schemas to the LLM, and maps tool names to executable code. The example CalculatorTool provides a schema and an execute() that evaluates expressions (tutorial uses eval() for simplicity and warns against it in production). The tutorial demonstrates why tools matter: the LLM produced a plausible but incorrect multiplication result, whereas delegating the math to the calculator returned the accurate 77,585.1801. Key implications: building from first principles teaches memory management (and context‑window limits), tool schema/dispatch patterns, safety caveats (no eval in prod), and how to instrument and debug agent reasoning before moving to higher‑level frameworks.