Learn Claude Code builds bash harness

The interesting part here isn’t that it uses bash, it’s that it turns “how do agent harnesses actually work?” into a teachable progression. That makes it useful both as a learning resource and as a stripped-down reference for teams prototyping local agents.

• –Starts with the smallest possible control loop, which makes the agent boundary easy to understand before any abstraction creep sets in
• –Adds planning, subagents, and skills in a way that mirrors the real scaling problems agentic coding systems hit
• –The later sections on context compaction, background tasks, and team protocols are the bits that matter once a toy agent becomes a daily tool
• –MIT licensing and local-run setup make it practical as a hands-on reference, not just a conceptual writeup