Build AI agents from scratch so frameworks finally make sense

What it does

This repository is a curriculum of fifteen runnable JavaScript examples that walk you from loading a local LLM all the way up to ReAct agents, memory systems, and embedding-based tool routing. It uses node-llama-cpp to keep inference on-machine, and every example includes a CODE.md for line-by-line breakdowns and a CONCEPT.md for the architectural idea. A companion website provides the mental-model map while the repo supplies the terrain you actually run.

The interesting bit

Rather than hiding function calling or reasoning loops behind library sugar, each example exposes the raw prompt construction, JSON schema negotiation, and token streaming so you can see the exact moment text generation becomes agency. The progression is deliberately pedagogical: you start with a basic prompt/response cycle and end with beam search over reasoning branches.

Key highlights

• Runs fully offline via node-llama-cpp; local model weights required (8 GB RAM minimum, 16 GB recommended).
• Fifteen progressive examples covering ReAct, AoT, Chain/Tree/Graph of Thought, persistent memory, typed error taxonomies, and embedding-based tool routing.
• Every example ships with line-by-line code explanations (CODE.md) and high-level pattern guides (CONCEPT.md).
• Includes a companion website for learning-path visualization and conceptual explanations, kept separate from the executable code.
• A Python sibling repository exists if you prefer that ecosystem.

Caveats

• Model binaries are not included; you must download and place them in a ./models/ directory before anything runs.
• One early example can call OpenAI endpoints, but the core learning path is entirely local.
• Local inference demands real hardware: the README lists 8 GB RAM as the floor and 16 GB as the practical minimum.

Verdict

Reach for this if you want to understand agent internals before adopting a framework, or if you are designing curriculum around local LLMs. Look elsewhere if you need a drop-in production framework today.