A minimal coding agent harness in Python, built to show how the pieces of a coding agent fit together.
This is the repo for the Neural Breakdown video on building Codign Agent from scratch. You can watch the full ~50 minute step-by-step walkthrough here. You can cycle through individual commits of this repo (other than the README edits) and see how the full project came to life.
coding-agent-star-patterns-edited-github.mp4
- Interactive terminal chat
- Tools for running shell commands, reading and writing files, and making targeted edits.
- Configurable model and OpenAI-compatible API endpoint.
- Tool permissions and shell sandboxing on macOS and Linux
- Skills loaded from project and user
.agents/skillsdirectories. - Subagents for exploring a codebase in a separate context window.
- Todo tracking for tasks with multiple steps.
- Saved chat sessions, with
/sessionsto reopen them and/rewindto go back in the conversation. - Automatic context compaction, plus
/compactto trigger it manually. - Git branch context and reminders when files change between turns.
If you find this helpful, consider supporting on Patreon — it hosts all code, projects, slides, and write-ups from the YouTube channel.
