Build the vehicle, not the driver: a nano Claude Code harness

What it does

This repository is a 20-chapter tutorial (with a legacy 12-lesson track) that walks through building a minimal agent harness in Python. It strips the concept down to a single loop: the LLM decides whether to use a tool, the harness executes it, and the cycle repeats. Each lesson layers on one real-world mechanism—tools, permissions, context compaction, subagents, task graphs—without ever touching the core loop.

The interesting bit

The project’s central argument is that most “agent frameworks” are delusional Rube Goldberg machines: they try to code intelligence with if-else branches and node graphs, when agency is actually learned during model training. It reframes the engineer’s job as building the vehicle (the harness) while letting the model remain the driver. The philosophical rant is backed by a concrete, runnable implementation that stays minimal.

Key highlights

• Teaches harness engineering, not prompt plumbing
• Builds a single persistent agent loop and layers on tools, sandboxing, and context management
• Covers advanced patterns like subagent spawning, async mailbox coordination, and worktree-isolated parallel execution
• Includes 20 narrative chapters with translations and runnable code.py files
• Explicitly rejects no-code workflow builders and procedural rule pipelines as “shell scripts with grandiose pretensions”

Caveats

• The repo currently maintains two overlapping tutorial tracks (root-level s01-s20 and a legacy 12-lesson version), which the README warns may confuse readers following old links
• It is explicitly a teaching repository, not a production drop-in replacement for Claude Code

Verdict

Worth a look if you are a developer who wants to understand why most agent frameworks are over-engineered and how to build the minimal infrastructure that actually lets an LLM work. Skip it if you are hunting for a batteries-included framework to drag and drop.