The best engineering frameworks are the ones you can explain in three sentences and build in a weekend.
The Summary
- A new open-source repo teaches how to build a Claude-like coding agent from scratch using just bash scripts, stripping away the mysticism around "agent harnesses"
- The companion guide details production patterns: claude.md files for context, skills libraries, subagents, plugins, and MCPs for using Claude as a daily coding driver
- Core thesis: agency comes from the model's training, not the orchestration code. The harness is just the vehicle for what the model already learned to do.
The Signal
The shareAI-lab repo cuts through agent product confusion with a clean distinction: agency lives in the model, infrastructure lives in the harness. When you spin up an AI coding assistant, the model already knows how to reason about code from billions of training examples. Your job is building the vehicle, not teaching it to drive.
The repo calls itself "nano claude code" and delivers on the promise. It's agent harness engineering reduced to first principles, using bash as the primary glue. No framework bloat. No 15-layer abstraction cake. Just the minimal scaffolding to let a capable model do its work.
"Agency was never bestowed by the surrounding code. It was learned during training."
The production implementation guide shows what this looks like when you run it daily. Five key patterns emerge:
- claude.md files: Project context documents that give the model the lay of the land
- Skills libraries: Reusable capabilities the agent can invoke
- Subagents: Specialized model instances for specific tasks
- Plugins: External tool integrations
- MCPs (Model Context Protocol): Structured ways to feed runtime context
The historical grounding matters here. The repo traces the lineage back to DeepMind's 2013 DQN, which learned to play Atari from raw pixels and scores alone. No game-specific code. No human feature engineering. Just a neural network that figured out what to do through training, then executed through a simple action interface.
That same principle scales to coding agents. The model learned to write code, debug, refactor, and architect from its training data. The harness just needs to handle I/O: read the codebase, execute tools, write files, return results. The 296 upvotes and 210 comments on Hacker News suggest developers are hungry for this clarity.
What makes this approach powerful is what it strips away. Most "agent frameworks" add layers of prompt orchestration, retrieval pipelines, and decision trees, implying the model needs help reasoning. This repo says the opposite: get out of the model's way. Give it context, give it tools, let it work.
The Implication
If you're building agent products, stop goldplating the harness. The model is already smarter than your orchestration logic. Your code should be thin infrastructure: authentication, file system access, API calls, output formatting. That's it.
For developers using Claude or similar models daily, the patterns here are immediately applicable. A claude.md file in your project root costs nothing and saves the model (and you) from re-explaining your architecture every session. Skills libraries turn one-off solutions into reusable capabilities. The rest is just bash and model calls.