betta-tech/harness-sdd

This repository is an example project that demonstrates a structured approach to using AI coding agents, specifically Claude Code, to develop software autonomously while keeping a human in control of important decisions. The actual application inside it is a simple command-line notes tool in Python, but that is deliberately trivial, the point is to show how the repository itself is organized to guide AI agents effectively. The core idea, which the project calls "Harness Engineering," is that a repository should be set up so that an AI agent can understand what to build, how to build it, and how to verify its own work, all from the files in the repository rather than from conversation history that gets lost. This is achieved through a structured workflow: before any code is written, the AI produces specification documents covering what a feature must do, what technical decisions were made, and a checklist of implementation steps. A human must approve these documents before the agent touches any code. After implementation, a separate reviewer agent checks that every requirement maps to an actual test. Multiple specialized sub-agents handle different roles, one for writing specs, one for implementing, one for reviewing, and they communicate by writing files to disk rather than passing messages through the chat, which makes the process resilient to context window limits and system restarts. Progress files are kept updated throughout so the state of the work is always visible. This is useful as a template or learning resource for developers who want to use AI agents for real development work with oversight and traceability built in. It is written in Python.