AgentHarness: build Claude Code from scratch
Learn harness engineering by building a Claude Code–style agent from scratch.
- Problem
- Developers want to learn how improving the harness—not the model—can boost coding agent performance, but systematic hands-on material is scarce. After taking apart Claude Code, some find that harness engineering brings significant gains, yet a from-scratch course is hard to find.
- Solution
- A hands-on course that teaches users to build a Claude Code–style coding agent harness from scratch. It analyzes cases like LangChain where the same model improved benchmark scores through harness optimization, and provides runnable code templates and experiment comparisons.
- Users
- Developers interested in AI coding agents, engineers wanting to improve agent performance, and technical team leads.
- MVP
- 5–8 core lessons: harness concepts, build a minimal coding agent from scratch, tool calling and sandboxing, context management, benchmark comparisons; runnable code repository; community Q&A.
- Revenue
- One-time course purchase at $49–$99, or annual subscription with updates; enterprise team licensing available.
- Why now
- The AI coding agent wave continues, and harness engineering is a proven direction but lacks systematic courses. An indie developer can quickly produce one based on their own teardown experience.