Coding agents can feel like magic: point one at a codebase, ask it to fix a bug, and it reads files, edits code, runs commands, and iterates until the task is done. But under the hood, the core ideas are surprisingly approachable.

In this live-coding talk, we’ll build a small coding agent from scratch and use it to demystify how tools like Claude Code, Cursor, or Codex work. We’ll look at the agent harness: the application layer that talks to the model, maintains context, exposes tools, interprets tool calls, and decides what to send back into the loop.

Along the way, we’ll cover the tool-calling loop, file and command access, context management, project-level configuration, reusable skills, and how more advanced systems coordinate subagents.

This is a practical, technical session for developers who want to understand what coding agents actually do, what tradeoffs they make, and how to design one that is useful, controllable, and debuggable.

Alternative title (if there's an issue with naming Claude Code explicitly): Let’s Build Our Own Coding Agent

This session is very much hands-on and can be delivered as a classical conference talk with live coding and/or live demonstration, or as a workshop in which attendees build their own coding agents.