Harness engineering focuses on improving agent output by shaping the environment around them, keeping the model and coding agent constant. This approach enhances external levers like context and tools, curating the surrounding environment. The goal is for the worker to recover intent, operate systems, respect authority, prove outcomes, and improve future runs.

A central purpose of the harness is to embed an organization's nonfunctional requirements—such as reliability, security, and performance—into code. This repository, "Make the Repository Teach the Agent," details how these requirements and decisions become retrievable context, examples, tools, and executable constraints. Harness engineering aims to make organizational judgment cumulative through iterative feedback loops, turning lessons from accepted work, failures, and user responses into context that shapes future actions.

Code serves as an agent's language for interacting with computers, enabling reliable domain outcomes without direct implementation review at the point of deployment. General model weights only represent a fraction of an organization's process data. Harness engineering makes crucial operational state, local ontology, quality standards, procedures, and authority relationships available to agents as context and tools. The repository provides guidance for agents, with AGENTS.md directing tasks and playbooks for practical application.