Coding agents like Claude Code and Codex can feel magical, but their core architecture is understandable and buildable. In this workshop, you’ll implement a compact coding agent from scratch and learn the engineering decisions that make it useful in practice.

We’ll build the full loop: model orchestration (frontier, open-source, or local), a tool harness for filesystem/terminal actions, lightweight memory, reusable skills, and a simple web interface for interaction and review. Along the way, we’ll focus on practical reliability and safety patterns: constrained execution, structured errors, idempotent actions, and clear approval boundaries for risky operations.

By the end, you’ll leave with a working reference implementation, a clear mental model of how modern coding agents operate, and a concrete path to evolve your prototype into a real developer tool.