14-lesson node.js course

Build the mind,
not the framework.

Most AI courses hand you LangChain and call it done. This one builds every agent pattern from raw model APIs — ReAct loops, memory, DAGs, tree search — so you actually understand what runs underneath.

14 lessons
~6 hours
Node.js
local LLMs
scroll to explore
↓
by the numbers
0
Lessons total
0h
Estimated duration
0
Learning stages
0%
Framework-free
0
Agent patterns built
the problem

Frameworks hide
what you need to know.

LangChain, LlamaIndex, AutoGPT — they all abstract away the one thing you actually need to understand: how the model thinks, calls tools, and loops.

✗Black-box abstractions that version-break every release
✗No mental model of what's actually running
✗Can't debug it because you never built it
✗Lock-in to one provider's design decisions
the solution

Build it once
from the ground up.

14 lessons, each a standalone working implementation. You write the ReAct loop. You wire the tool caller. You build the memory system. Then you own it.

✓Node.js + raw model APIs — nothing hidden
✓Local LLMs (llama.cpp) + hosted APIs (OpenAI)
✓Every pattern is 100–200 lines, no fluff
✓Transfer the mental model to any provider or framework
what you'll build

8 production-grade agent patterns

Function Calling
JSON-schema tool definitions the model calls by name — the building block of all agents.
ReAct Loop
Reason → act → observe, iterated until the model decides it has a final answer.
Memory System
Persist facts between sessions with keyword retrieval. Primer for vector embeddings.
DAG Executor
Decompose a task into atomic ops, resolve dependencies, run in parallel where possible.
Tree of Thought
Generate N reasoning branches, score deterministically, keep the best — beam search for LLMs.
Graph of Thought
Parallel extraction + conflict resolution before generation. For multi-source synthesis.
Chain of Thought
Sequential reasoning phases: facts → signals → policy → decision. Produces auditable traces.
Error Resilience
Typed error taxonomy, exponential backoff with jitter, graceful degradation per failure mode.
learning path

Three stages.
14 lessons. Zero fluff.

built-in ai tutor

An expert at your side for every lesson.

Each lesson page ships with a context-aware AI tutor powered by Mistral. The lesson concept and key patterns are injected into every conversation so answers stay grounded in exactly what you're learning.

Asks questions? The tutor knows the lesson cold.
Surfaces connections to adjacent lessons automatically.
Production use cases — not just toy examples.
API key stays server-side. Never in the browser.
AI tutorreact agent
Why does ReAct need an observation loop?
A single tool call can't solve multi-step problems — the model needs to see the result of each action before deciding the next one. The observation feeds back into context so the model can self-correct.
How does this connect to AoT?
ReAct discovers its plan as it goes; AoT plans the entire dependency graph upfront before executing. Use ReAct when structure is unknown, AoT when it can be enumerated.
Ask anything…
ready to build?

Start with Lesson 1.
It takes 15 minutes.

Load a local model, run your first inference, and understand exactly what the rest of the course builds on.

logoagent
14 lessons · 3 stages · MIT license