LLM Agents From Scratch

The companion library for Build a Multi-Agent System — With MCP and A2A (Manning). Learn how LLM agents work by building one yourself, from first principles, step by step.

Available now through Manning's Early Access Program (MEAP) — buy today and get each chapter as it's completed. Buy the Book →

About

Multi-agent systems and the LLM agents that power them are among the most discussed topics in AI today. There are already many capable frameworks out there — the goal of this book isn't to replace them, but to help you deeply understand how they work by having you build one yourself, from scratch.

All the code lives in the book's own hand-rolled agent framework, primarily designed for educational purposes rather than production deployment. It will give you the foundation to work more confidently with any other LLM agent framework of your choosing, or even to build your own specialised solutions.

From the Book

Each chapter builds on the last, progressively deepening your understanding from core concepts to full multi-agent systems.

Part 1 — Build Your First LLM Agent

Part 2 — Enhance Your LLM Agent

Part 3 — Building Multi-Agent Systems

Capstone Projects

Capstones are larger, end-to-end projects that pull together what you have built in the book and apply it to something closer to a real-world system.

Running the Notebooks

The chapter notebooks use Qwen3:14B as the default model via Ollama. Minimum hardware: ~16 GB RAM with a discrete GPU, or any machine with a dedicated GPU that can comfortably run a 14B model. You can swap in a smaller model (e.g. qwen3:4b), but note that smaller models may not follow tool-use instructions reliably — especially in later chapters.

Option A — Run locally

Install Ollama and pull the model:

ollama pull qwen3:14b

Then launch Jupyter from the project root:

uv run --with jupyter jupyter lab

Option B — Lightning AI (recommended if you don't have a GPU)

Lightning AI offers a free tier with ~22 GPU compute hours/month — enough for several sessions on an L4 GPU.

1. Create a free Lightning AI account if you don't have one
2. Visit the book template: lightning.ai/nerdai/templates/build-a-multi-agent-system-from-scratch
3. Click Clone to copy it to your account
4. Make sure you have the latest version of the template
5. Launch the Studio with an L4 GPU
6. Open the chapter notebook you want to run and execute the cells — Ollama and the required model are already set up in the template

Getting Started

Prerequisites

• Python 3.10+
• Ollama running locally (used as the default LLM backend)
• uv for dependency management

Installation

Clone the repository:

# SSH
git clone git@github.com:nerdai/llm-agents-from-scratch.git

# HTTPS
git clone https://github.com/nerdai/llm-agents-from-scratch.git

cd llm-agents-from-scratch

Install dependencies:

uv sync --all-extras --dev

Quick Start

from llm_agents_from_scratch.llms import OllamaLLM
from llm_agents_from_scratch.agent import LLMAgentBuilder
from llm_agents_from_scratch.tools import SimpleFunctionTool

def add(a: int, b: int) -> int:
    return a + b

llm = OllamaLLM(model="llama3.2")
tool = SimpleFunctionTool(fn=add)

agent = (
    LLMAgentBuilder()
    .with_llm(llm)
    .with_tools([tool])
    .build()
)

result = await agent.run("What is 3 + 5?")
print(result)

Development

# Run all tests
make test

# Lint and format
make lint
make format

# Coverage report
make coverage-report

See CLAUDE.md for full development guidance.

Contributing

Bug reports, feature requests, and community project submissions are welcome. See CONTRIBUTING.md for details.

• Found a bug? Open an issue
• Built something cool? Share it on GitHub Discussions

License

Apache 2.0 — see LICENSE for details.