Repolode
Submit
decodingai-magazine

decodingai-magazine/building-a-coding-agent-from-scratch-course

A free, open-source course that teaches harness engineering by building a Claude Code-style coding agent from scratch in Python.

Open-source alternative to:Claude Code

View on GitHub
Stars
1,129
Forks
267
Open issues
6
Last commit
14 days ago

What is it?

This course walks through building a coding agent called decode, from a bare agent loop up to a fleet of cloud agents. The core is a headless harness: the tool-calling loop, tools for file I/O, bash, grep and edits, a permission layer, and a sandbox. The harness is wrapped in two interfaces, an interactive terminal UI and remote runs on Modal serverless functions that can be triggered by CLI, webhook or cron.

The course also covers context engineering (memory files, skills, compaction, LSP diagnostics), subagents that fan out in parallel, and evaluation. Evals use benchmarks and regression suites traced in Opik, while session recording and replay are handled through Kitaru. Model providers include Modal-hosted open weights, OpenRouter and Gemini.

The material is delivered as 8 written lessons and 6 videos, with setup guides in the repository. Code is written in Python on Pydantic AI, and the components are built by hand rather than taken from frameworks.

Key features

  • 01Builds a tool-calling agent loop with read, edit, bash, grep and other tools.
  • 02Runs the agent in a terminal UI or headless on Modal serverless functions.
  • 03Executes tool commands in local Docker or remote Modal sandboxes.
  • 04Includes a permission layer with allow, ask and deny rules.
  • 05Covers context engineering: memory files, skills and auto-compaction.
  • 06Provides eval tooling with benchmarks and regression tests traced in Opik.

Use cases

  • Learning how production coding agents like Claude Code are structured internally.
  • Building a custom agent for a domain other than coding, using the same harness patterns.
  • Running open-weight models served on Modal as the backend for an agent.
  • Practicing agent evaluation with benchmarks and regression suites.

Quick start

  1. 1

    Clone the repository

    git clone https://github.com/decodingai-magazine/building-a-coding-agent-from-scratch-course.git
  2. 2

    Install dependencies from inside the project folder

    make install
  3. 3

    Copy the example env file, then set your LLM API key in it

    cp .env.example .env
  4. 4

    Start the decode agent in the terminal

    uv run decode

Things to know

  • Running the agent requires an LLM API key from Gemini, OpenRouter, or Modal.
  • Some features depend on third-party services such as Modal, Opik, and Kitaru.

In collections

Claude Code tools, skills & pluginsOpen-source AI agent frameworks & appsAI coding assistants & agents

Topics

#ai-agent#coding-agent#llm#agent-evals#sandboxing#context-engineering#python#tutorial#ai-coding-assistant#claude-code

FAQ

What is building-a-coding-agent-from-scratch-course?
A free, open-source course that teaches harness engineering by building a Claude Code-style coding agent from scratch in Python. This course walks through building a coding agent called decode, from a bare agent loop up to a fleet of cloud agents. The core is a headless harness: the tool-calling loop, tools for file I/O, bash, grep and edits, a permission layer, and a sandbox. The harness is wrapped in two interfaces, an interactive terminal UI and remote runs on Modal serverless functions that can be triggered by CLI, webhook or cron.
Is building-a-coding-agent-from-scratch-course free to use? What license does it have?
building-a-coding-agent-from-scratch-course is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is building-a-coding-agent-from-scratch-course still maintained?
Actively maintained. The last commit to building-a-coding-agent-from-scratch-course was on Oct 24, 2026.
What can I use building-a-coding-agent-from-scratch-course for?
Learning how production coding agents like Claude Code are structured internally. Building a custom agent for a domain other than coding, using the same harness patterns. Running open-weight models served on Modal as the backend for an agent. Practicing agent evaluation with benchmarks and regression suites.
How do I install building-a-coding-agent-from-scratch-course?
Clone the repository: git clone https://github.com/decodingai-magazine/building-a-coding-agent-from-scratch-course.git Install dependencies from inside the project folder: make install Copy the example env file, then set your LLM API key in it: cp .env.example .env Start the decode agent in the terminal: uv run decode
Where can I get building-a-coding-agent-from-scratch-course?
The source code and installation instructions are on GitHub: https://github.com/decodingai-magazine/building-a-coding-agent-from-scratch-course

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects