# AI Agents from Scratch

AI Agents from Scratch is a 12-lesson, local-first course that builds one AI agent step by step from a single local LLM call, with no frameworks and no cloud APIs, so a developer can see exactly how the loop, tools, memory and planning underneath an agent actually work.

AI Agents from Scratch is a local-first introduction to how AI agents work, building one agent step by step from a single local LLM call with no frameworks, cloud APIs or hidden reasoning. Its stated view is that agents are not personalities but loops, state and constraints, and every mechanism in the repository is meant to be readable in the code itself.

Across 12 lessons it adds, in order, plain text exchange, system prompts, structured JSON output, decision routing, external tools, a full observe-decide-act agent loop, short and long-term memory, planning represented as data, safe atomic actions, dependency-graph planning, regression evals, and runtime telemetry. It targets developers who already know how to code but feel lost with agent frameworks, and explicitly is not aimed at people who want the fastest demo or a ready-made SaaS starter kit.

Website: https://github.com/pguso/agents-from-scratch