AI Agents From Scratch

Learn to build AI agents locally without frameworks. Understand what happens under the hood before using production frameworks.

Purpose

This repository teaches you to build AI agents from first principles using local LLMs and node-llama-cpp. By working through these examples, you'll understand:

• How LLMs work at a fundamental level
• What agents really are (LLM + tools + patterns)
• How different agent architectures function
• Why frameworks make certain design choices

A Python version of this tutorial is available here: https://github.com/pguso/agents-from-scratch

Philosophy: Learn by building. Understand deeply, then use frameworks wisely.

Companion Website

This repository now has a matching companion website:

https://agentsfromscratch.com

The website is not a replacement for this repo, but a conceptual companion that:

• Explains why each example exists
• Visualizes the learning path from raw LLM calls to full agents
• Separates code, explanations, and core concepts
• Helps you understand agent architectures before using frameworks

Recommended workflow:

• Use GitHub for running, modifying, and studying the code
• Use the website for mental models, explanations, and progression

Think of the site as the map and this repo as the terrain.

Agent Fundamentals - From LLMs to ReAct

Prerequisites

• Node.js 18+
• At least 8GB RAM (16GB recommended)
• Download models and place in ./models/ folder, details in DOWNLOAD.md

Installation

npm install

Run Examples

node intro/intro.js
node simple-agent/simple-agent.js
node react-agent/react-agent.js

Learning Path

Follow these examples in order to build understanding progressively:

1. Introduction - Basic LLM Interaction

intro/ | Code | Code Explanation | Concepts

What you'll learn:

• Loading and running a local LLM
• Basic prompt/response cycle

Key concepts: Model loading, context, inference pipeline, token generation

---

2. (Optional) OpenAI Intro - Using Proprietary Models

openai-intro/ | Code | Code Explanation | Concepts

What you'll learn:

• How to call hosted LLMs (like GPT-4)
• Temperature Control
• Token Usage

Key concepts: Inference endpoints, network latency, cost vs control, data privacy, vendor dependence

---

3. Translation - System Prompts & Specialization

translation/ | Code | Code Explanation | Concepts

What you'll learn:

• Using system prompts to specialize agents
• Output format control
• Role-based behavior
• Chat wrappers for different models

Key concepts: System prompts, agent specialization, behavioral constraints, prompt engineering

---

4. Think - Reasoning & Problem Solving

think/ | Code | Code Explanation | Concepts

What you'll learn:

• Configuring LLMs for logical reasoning
• Complex quantitative problems
• Limitations of pure LLM reasoning
• When to use external tools

Key concepts: Reasoning agents, problem decomposition, cognitive tasks, reasoning limitations

---

5. Batch - Parallel Processing

batch/ | Code | Code Explanation | Concepts

What you'll learn:

• Processing multiple requests concurrently
• Context sequences for parallelism
• GPU batch processing
• Performance optimization

Key concepts: Parallel execution, sequences, batch size, throughput optimization

---

6. Coding - Streaming & Response Control

coding/ | Code | Code Explanation | Concepts

What you'll learn:

• Real-time streaming responses
• Token limits and budget management
• Progressive output display
• User experience optimization

Key concepts: Streaming, token-by-token generation, response control, real-time feedback

---

7. Simple Agent - Function Calling (Tools)

simple-agent/ | Code | Code Explanation | Concepts

What you'll learn:

• Function calling / tool use fundamentals
• Defining tools the LLM can use
• JSON Schema for parameters
• How LLMs decide when to use tools

Key concepts: Function calling, tool definitions, agent decision making, action-taking

This is where text generation becomes agency!

---

8. Simple Agent with Memory - Persistent State

simple-agent-with-memory/ | Code | Code Explanation | Concepts

What you'll learn:

• Persisting information across sessions
• Long-term memory management
• Facts and preferences storage
• Memory retrieval strategies

Key concepts: Persistent memory, state management, memory systems, context augmentation

---

9. ReAct Agent - Reasoning + Acting

react-agent/ | Code | Code Explanation | Concepts

What you'll learn:

• ReAct pattern (Reason → Act → Observe)
• Iterative problem solving
• Step-by-step tool use
• Self-correction loops

Key concepts: ReAct pattern, iterative reasoning, observation-action cycles, multi-step agents

This is the foundation of modern agent frameworks!

---

10. AoT Agent - Atom of Thought Planning

aot-agent/ | Code | Code Explanation | Concepts

What you'll learn:

• Atom of Thought methodology
• Atomic planning for multi-step computations
• Dependency management between operations
• Structured JSON output for reasoning plans
• Deterministic execution of plans

Key concepts: AoT planning, atomic operations, dependency resolution, plan validation, structured reasoning

---

11. Error Handling - Resilience for LLM + Tools

error-handling/ | Code | Code Explanation | Concepts

What you'll learn:

• Typed error taxonomy (validation, LLM, tools, workflow) with stable codes
• Timeouts, retries with backoff/jitter, and classifying transient failures
• Graceful degradation when the LLM path fails (deterministic tool fallback)
• Orchestration-level errors (AgentWorkflowError) and correlation ids for support

Key concepts: Error taxonomy, retry policies, timeouts, fallbacks, degraded mode, observability, user-safe messaging

---

12. Tree of Thought - Search over reasoning branches

tree-of-thought/ | Code | Code Explanation | Concepts

What you'll learn:

• Generating multiple candidate next actions from the same partial plan
• Ranking and pruning branches with a deterministic score in code
• Running a compact beam search loop with inspectable kept/pruned decisions
• Verifying the winning path with explicit sanity checks

Key concepts: Tree of Thought, beam search, branch pruning, verifiable objectives, search controllers

---

13. Graph of Thought - DAG merge for multi-source outputs

graph-of-thought/ | Code | Code Explanation | Concepts

What you'll learn:

• Modeling reasoning as a DAG: parallel source extracts → merge rules → final draft
• Resolving conflicts explicitly before generation (must_include, must_avoid, conflict_notes)
• Adding deterministic merge and draft compliance checks
• Running independent nodes in parallel to reduce latency

Key concepts: Graph of Thought, DAG orchestration, multi-source fusion, merge-before-generate, policy reconciliation

Decision guide: use ToT when you need to search competing paths; use GoT when you need to combine multiple sources into one consistent policy. Compare both in:

• ToT concept
• GoT concept

---

14. Chain of Thought - Auditable stepwise decisioning

chain-of-thought/ | Code | Code Explanation | Concepts

What you'll learn:

• Splitting a high-stakes decision into explicit reasoning phases
• Preventing early bias with a facts-only extraction step
• Balancing fraud signals with legitimacy evidence before policy application
• Producing an auditable final decision with customer-safe and internal outputs

Key concepts: Chain of Thought, structured reasoning traces, policy-constrained decisions, explainability, review-ready workflows

---

15. Tool routing (embeddings) - Narrow the tool catalog per request

tool-routing-embeddings/ | Code | Code Explanation | Concepts

**What you