pguso/ai-agents-from-scratch — reverse-engineered prompt

Build me a JavaScript learning repo that teaches how AI agents work from scratch, using local language models instead of hiding everything behind a big framework.

I want it to feel like a hands on course. Start with a simple local model call, then add examples for prompting, translation, reasoning, streaming answers, batching, tool use, memory, and a ReAct style agent that can think, act, observe, and try again. Add later lessons for planning, error handling, and tree of thought style search if that fits.

Each lesson should have runnable code, a plain English explanation of the code, and a separate concept note explaining why that pattern matters. Include setup instructions, model download instructions, an example env file, and a clear learning path in the README. Keep the code simple and readable so someone can understand what is happening under the hood. Use Node.js, and look up current local LLM docs online if you need to.