Summary

This repository is a learning guide and archive for Harness Engineering, an OpenAI-proposed paradigm where engineers design AI agent environments, intentions, and feedback loops instead of writing code. It targets engineers and researchers interested in AI engineering, offering a structured path from concept to practice for reliable AI-driven task completion.

How It Works

Harness Engineering shifts engineer output from code to a "constraint system" including AGENTS.md navigation, architectural rules, custom linters, and feedback loops. Key principles: repository as source of truth, AGENTS.md as a "map" for progressive disclosure, mechanical enforcement via linters with repair instructions, optimizing for "agent readability" with stable tech, short PR lifecycles due to high agent throughput, and managing technical debt via automated scanning.

Quick Start & Requirements

The project outlines a learning path (Phases 1-5) covering conceptual study, critical thinking, practical experimentation with AI agents, feedback documentation, and work production. It's a methodology, not a direct tool installation. While the guide has no explicit prerequisites, practical application implies AI agent requirements. Links to concept notes are provided.

Highlighted Details

• Efficiency: Estimated development time reduced to ~1/10th of manual coding.
• Throughput: Expanded team (3-7 members) averages 3.5 PRs/person/day; runs exceed 6 hours.
• Repository Structure: Nested AGENTS.md files enable progressive disclosure ("map not manual").
• Core Concepts: Six principles: Repo-as-Source-of-Truth, Map-not-Manual, Mechanical Enforcement, Agent Readability, Throughput-Driven Merging, Entropy Management.

Maintenance & Community

The project is actively developing, with a team growing from 3 to 7 members over five months and ~1,500 PRs. Contributions are welcomed via Issues/PRs for concept notes, insights, practical cases, and resources. Contact available via GitHub, X (Twitter), and Telegram.

Licensing & Compatibility

Released under the MIT license, permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

The README cautions that shared experiences require context-specific adaptation. As an evolving "learning project" and "archive," it's a developing methodology, not a finalized framework. Specific AI agent implementation details are likely in practice/works sections.