Build Your Own Coding Agent from Scratch

Matt Wynne & Steven Diamante 2-Day Workshop
Mon-Tue, Sept 21-22

Build an AI agent from scratch, without any frameworks, so you understand what's actually happening under the hood.

You use AI agents every day: Claude Code, Codex, Cursor. But do you know what's actually happening under the hood? It turns out, the core of every AI agent is surprisingly simple: a while loop, a language model, and some tools.

In this hands-on workshop, we'll build an AI agent from scratch without any frameworks. We'll start from a bare frontier model call to an LLM and incrementally add capabilities. At each step, we'll hit a real limitation, understand why it exists, and solve it. You'll watch a basic chatbot evolve into a sophisticated agent with tools, conversation history, and guardrails.

What you will learn

By the end of day one, you'll be able to:

• Explain the agentic loop that powers LLM agents
• Identify the core components of AI agents
• Build your own simple agent and extend it with custom tools
• Explain context windows and how to manage agent context

On day two we'll explore deeper topics:

• Deterministic vs non-deterministic capabilities of agents: when and where each is appropriate
• Using evals to test the non-deterministic behaviour of your agent
• Adding domain-specific tools to manage ludic constraints

We'll attempt a refactoring challenge, and pit our agents against each other in a fun competition to end the workshop.

Who this workshop is for

• Developers and engineers who want to understand AI agents from first principles, not just use them as black boxes
• Tech leads and architects evaluating where agents fit in their systems and where they don't
• Anyone curious about what's actually happening when Claude Code, Codex, or Cursor runs

Agenda

Day 1: From LLM call to working agent

• The agentic loop: while loop, model, tools
• The core components of an AI agent
• Building a simple agent and extending it with custom tools
• Context windows and managing agent context

Day 2: Making agents useful and reliable

• Deterministic vs non-deterministic capabilities, and when each is appropriate
• Using evals to test non-deterministic behaviour
• Adding domain-specific tools to manage constraints
• A refactoring challenge, and a competition between the agents you've built

How it works

This is a hands-on workshop. You'll build code, hit real limitations, and solve them. Demos and discussion frame each step, but most of the time you're at the keyboard. We'll be using TypeScript to build the agent, with Java samples for the refactoring exercises. You don't need to be especially familiar with either language.

The Trainers

Matt Wynne

Matt was lucky enough to start programming in the mid 1980s when his dad brought home a BBC Micro from work.

He's been making a living as a programmer since the late 1990s, and started trying to sneak agile practices into his work ever since he discovered the C2 Wiki in 2001. He was once fired for doing Test-Driven Development (TDD), because he was "going too slow".

In 2008 he joined the fledgling Cucumber open source project, and in 2011 the first edition of The Cucumber Book was published, co-written with Aslak Hellesøy, the creator of Cucumber.

In 2013 Matt co-founded Cucumber Limited together with Aslak and Julien Biezmans. They grew this consulting and training business to around $1.5m turnover, training thousands of people around the world in Cucumber and Behaviour-Driven Development (BDD), adding Seb Rose and Steve Tooke then Dr. Sal Freudenberg as partners.

Matt developed the practice of Example Mapping whilst training a team in St. Louis, Missouri.

Matt acted as Cucumber Ltd's CEO for most of it's life, handing over to Sal in the autumn of 2018. In 2019 Cucumber was acquired by SmartBear. Matt worked for them as full-time lead of the Cucumber Open source project until February 2023. One of his initiatives was finding ways to tackle the structural sexism and racism in open source.

Originally from northwest England, he now lives in the mountain town of Nelson, BC in Canada on the unceeded territory of the Ktunaxa Nation, Sinixt Nation, and Syilx Peoples with his wife, two kids, a dog and a couple of cats. He enjoys back-country skiing, trail running and woodworking.

Steven Diamante

Steven Diamante is a technical coach who helps software teams get real value from AI coding agents. Most teams use AI to generate code fast, but then spend hours debugging. Steven teaches teams to use TDD and strict quality gates to keep AI agents steerable, ensuring they build the right thing in small, controllable steps. Through hands-on Samman Coaching, he helps teams transform how they work with AI while maintaining technical excellence.

Outside of coaching, he enjoys playing guitar in his local church, reading a non-fiction book, or hanging out with his wife and two kids.

FAQ

What level of experience do I need?

This is an intermediate-level workshop. You should be comfortable reading and writing code in at least one mainstream language. You don't need to be a TypeScript or Java expert; the exercises are structured so you can follow along regardless.

Should I bring my laptop?

Yes. You will be writing code throughout both days.

Do I need an API key?

Yes. You'll need an API key from a frontier model provider (Anthropic, OpenAI, or Google Gemini). You're responsible for any API usage charges incurred during the workshop, which are typically small.

Is this a workshop about using AI tools to write code?

No. This workshop is about building the kind of tool you use every day, so you understand it from the inside. The skills you'll learn transfer to evaluating, customizing, and extending the agents you use in your day job.

Workshop Details

Time: 9:00 am to 5:00 pm each day

Max participants: 30

Prerequisites: Comfort reading and writing code; a laptop; an API key for a frontier model provider (Anthropic, OpenAI, or Google Gemini)

Included: Coffee breaks and lunch with the trainers and fellow attendees are included in the workshop ticket.