Loopath

Learn how coding agents actually work — by building the loop yourself, inside your agent.

_{▶︎ Watch the full 2-min intro on YouTube}

---

Most agent tutorials stop at prompt → model → answer. That's not the harness.

Loopath is an interactive course that lives inside Codex or Claude Code and teaches loop engineering — the runtime system around a coding agent — by having you build a tiny, complete one from scratch. It's nanoGPT, but for the agent loop.

It teaches one small step at a time, in your language, with a short narrated clip for every step. Not a doc you read — a course your agent runs.

Install

# Claude Code
git clone --depth 1 https://github.com/encircleacity2/loopath ~/.claude/skills/loopath
# Codex — same repo, swap the path:  ~/.codex/skills/loopath

Restart your agent, then say:

Use $loopath to start the course.

The skill auto-detects whether you're writing in 中文 or English and teaches in that language — matching video and content.

The loop you'll build

context → action → policy → tool → observation → trace → eval → ↺

• Why a harness is a runtime system, not an API wrapper
• Structured action schemas, tool registries, and policy gates
• Traces, evals, loop engineering, and self-repair
• 14 episodes → one small, runnable, inspectable agent harness you can explain in an interview

How it works

• 🎬 One step at a time — each with a short narrated clip (中文 / English)
• 🌐 Fully bilingual — auto-routes language for video and explanations
• 🧪 Hands-on vibe-coding lab — you write the implementation request, the agent builds, verifies, and grades it
• ❓ Quiz with grading to check you actually understood it

Getting started in 20 seconds

1. git clone … ~/.claude/skills/loopath (above)
2. Restart your agent
3. Say "Use $loopath to start the course"
4. Watch the intro, then learn the loop one step at a time

Videos stream from external hosting by default, so the repo stays lean and git clone is fast. In Codex, the skill can also ask whether to cache all videos locally first, then embed each step as a playable video during the tutorial.

中文快速开始

Loopath 是一个在 Codex / Claude Code 里运行的中英双语互动课程，带你从零搭一个最小但完整的 coding-agent harness，真正理解 agent 背后的 loop engineering。

安装（Claude Code，Codex 把路径换成 ~/.codex/skills/loopath）：

git clone --depth 1 https://github.com/encircleacity2/loopath ~/.claude/skills/loopath

重启 agent，然后说：用 $loopath 开始 Loopath 课程。skill 会根据你的语言自动选择中文或英文，视频和讲解都跟着切换。一次一个小课题，每个 step 配有旁白短视频，最后用 quiz 检验理解。导论视频：English (YouTube) · 中文。

How the skill is built & course materials

Loopath is a data-driven skill — no course engine script to run. The agent is the course runtime:

• course/episodes.json — the bilingual source of truth: all 14 episodes' titles, theses, learning objectives, step-by-step teaching beats, lab deliverables/verification, and a per-episode quiz bank with reference answers. Each episode's lab points at its on-disk templates.
• course/labs/episode-NN/ — ready-to-write, English lab templates for every episode (modules, tests, demo repos, prompts, docs), laid out exactly as they belong in the student's project. Integration edits to earlier files ship as *.snippet.md. Every .py template is py_compile-clean and the core suite passes (pytest).
• SKILL.md — instructions that tell the agent to read that data, teach one step at a time, run labs as a vibe-coding loop (learner prompt → agent implementation → verification → prompt feedback), grade quizzes against the reference answers + rubric, and resolve videos from video_sources.json.
• Full course draft — the complete long-form course (full code listings + design discussion) the agent pulls from for depth, plus references/ for Episode 1.

Earlier versions shelled out to a scripts/loopath.py engine. That's gone: executing a freshly-cloned third-party script gets blocked by agent safety classifiers on first install, and everything it did (rendering cards, grading, scaffolding the lab) is something the model does better directly. Content lives in data; the model is the runtime.

Videos stream from external hosting by default, configured in video_sources.json. In Codex, the skill can offer to cache the full course locally with scripts/cache_videos.py; cached videos live under media/local/ and are ignored by Git.

• Bilingual step clips (Episode 1: 14 steps; Episodes 2–14: 5 steps each, zh+en, narrated) are hosted on GitHub Releases (clips-v1) and resolved via the clip_url_template. To switch to your own CDN (R2/TOS/etc.), edit that file.
• Intro: English (release asset) · 中文 (release asset)

License

MIT — share it, remix it, teach with it.