Awesome Harness Engineering

A curated, community-maintained list (an 'awesome list') of resources for harness engineering — the practice of designing everything that surrounds an AI agent but isn't the agent's brain itself.

The project argues that what makes an AI agent reliable on real tasks is often not the model, but the harness: the context delivered to it, the tools it can call, the planning and verification loops it runs in, its memory systems, its sandboxes. This list collects the best existing material on that discipline — essays, academic papers, engineering blog posts, reference implementations, and templates.

What you'll find

The README is organized into themed sections. The two largest are:

• Foundations — a long, annotated reading list of essays and papers that define the field. Contributors include OpenAI, Anthropic, Google, Meta, Microsoft, Red Hat, Martin Fowler, LangChain, and academic researchers. Each entry comes with a dense one-line summary written by the list maintainers.
• Design Primitives — 12 subcategories covering the building blocks of a harness: the agent loop, planning/task decomposition, context delivery and compaction, tool design, skills and MCP, permissions and authorization, memory and state, task runners and orchestration, verification and CI integration, observability and tracing, debugging and developer experience, and human-in-the-loop.

Other sections cover reference implementations (tutorials, meta-harnesses, demo harnesses), security/sandboxing, evals and verification, ready-to-use templates, production infrastructure, and links to related awesome lists.

Key claims embedded in the curation

The maintainers' editorial stance, repeated throughout the entries, is that the harness, not the prompt, is the real reliability surface for agents — and that as models improve, good harnesses are designed with the expectation that their own components will eventually become unnecessary.

Format and community signals

• License: CC0 (public domain dedication)
• Marked with the badge, meaning it follows the conventions of the awesome list ecosystem
• Includes a banner image, GitHub stars/forks/last-commit badges, and a link to join a community forum (linux.do)
• The README provides translation links to 9 languages via an external service (zdoc.app)
• A short Contributing section (not detailed in the README excerpt) invites community participation

Practical takeaway

If you build or operate AI agents and want to get better at the engineering around them — context management, tool interfaces, verification loops, permissions — this project is a single, well-filtered starting point to the best existing writing and code on the topic.