Agent Harnesses: Anatomy, Build, Frameworks, Benchmarks
Curated by @tripp
Updated Oct 2, 2026
Take apart Claude Code, Codex CLI, OpenCode, Pi and DeepSeek Harness on one 10-point rubric, build your own coding-agent harness in Python on OpenRouter, rebuild it on LangGraph, the OpenAI and Claude Agent SDKs and more, read the leading benchmarks critically, and ship it on a Terminal-Bench 4.0 slice through Harbor. 32 lessons, BYOK.
Learning is better together
Bring a question from this course or share what you made with other learners.
Quality
Scored on six categories — citations, structure, content richness, retrieval-readiness, scaffolding, and writing quality.
How densely lessons cite external sources, and how varied those sources are.
Module balance and reasonable lesson sizing.
Code examples, diagrams, and math — the multimodal payload.
How well-sectioned lessons are for search, citations, and deep links.
Models used
Models the creator uses in this course — chat, study guides, flashcards, ingest. Updated when the listing is republished.
Syllabus
Folder tree from your library — expand modules to see lessons.
Changelog
- v0.22026-10-02
Added "How this course was made" to the syllabus: a disclosure card for the course itself. Built in Claude (Cowork, Agent SDK harness) on claude-opus-5-5 with 73 subagents; ~304M input-side tokens (mostly cache reads) and 2.88M output tokens; 27.8 h wall clock, ~6 h active.
- v0.12026-10-02
First release: 9 modules, 32 lessons. Teardowns of the popular coding-agent harnesses (incl. Pi and DeepSeek Harness), tinyharness build with tested companion code, framework ports, the 2026 benchmark landscape, eval/RL harnesses on Harbor, and a Terminal-Bench capstone.
Attribution
LearnOS is a learning workspace, not a content host. This listing curates external resources; original creators retain ownership. Learners may need to obtain materials separately where noted.