ai-agents-from-scratch

AI Agents From Scratch is a code-first learning series that teaches how to build AI agents locally and without frameworks, using local LLMs and node-llama-cpp. The stated aim is to understand what happens under the hood before using production frameworks: how LLMs work at a basic level, what an agent really is (an LLM plus tools plus patterns), how different agent architectures function and why frameworks make the choices they do. The README links a long-form interactive version at agentsfromscratch.com for those who prefer diagrams and conceptual explanations.

Phase 1 takes you from LLMs to a ReAct agent through numbered examples, each with runnable code, a code explanation and a concepts page: an introduction to basic LLM interaction, an optional OpenAI example (temperature and token use), translation with system prompts, and further steps up to the ReAct agent. You need Node.js 18+, at least 8 GB of RAM (16 GB recommended) and downloaded model files placed in ./models/, then `npm install` and run the examples with node. A next phase is described as re-implementing core ideas of LangChain and LangGraph in plain JavaScript with local models, and related projects cover AI product development and a Python version. GitHub shows no licence.