Building a Coding Agent From Scratch
An open-source course that builds a Claude Code-style coding agent from scratch in Python, covering the tool-calling loop, permissions, sandboxing, memory, session recording and a subagent fan-out.
This course teaches how to build a coding agent's harness, the part the authors argue matters more than the underlying model, through 8 articles and 4 videos built around one codebase, produced by Decoding AI with Modal, Opik and Kitaru. The finished agent, decode, has a roughly 20-line tool-calling core; everything else, tools, skills, the permission layer, sandboxing, a steering queue, memory, compaction, session recording and replay, remote execution and subagent fan-out, is what the course builds.
After cloning the repository and running an install script, the finished agent can be tried immediately through demo skills accessible with a /demo- command, including generating a playable terminal game, building a live GitHub-data dashboard, and extracting an interactive knowledge graph from web articles. The agent runs as one headless core with two front ends: an interactive terminal UI and parallel Modal serverless functions triggered by CLI, webhook or cron.