Summary

This repository curates resources, patterns, and templates for "Harness Engineering," the discipline of designing the scaffolding around AI agents. It addresses the critical need for robust infrastructure—context delivery, tool interfaces, planning, verification, memory, and sandboxes—to ensure AI agents succeed on real-world tasks. Aimed at engineers and researchers, it provides a structured approach to building reliable AI agent systems, focusing on the surrounding architecture rather than the core models.

How It Works

The project is a curated collection of links to canonical essays, engineering blog posts, research papers, and reference implementations. It categorizes resources by harness engineering problem areas like Agent Loops, Planning, Context Delivery, Tool Design, Memory, and Verification. The core approach is to synthesize best practices and architectural patterns from leading AI labs (OpenAI, Anthropic, Google, Meta) and research, emphasizing that harness components are designed to compensate for current model limitations and will evolve as models improve.

Quick Start & Requirements

This is a curated list of resources, not a runnable software project. It serves as a knowledge base and reference guide.

Highlighted Details

• Extensive coverage of foundational concepts from OpenAI, Anthropic, and Martin Fowler, defining harness engineering.
• Detailed exploration of core components: Agent Loops (ReAct, LangGraph), Planning (Plan-and-Execute, LATS), Context Delivery (LLMLingua, RAG), Tool Design (MCP, Function Calling), Memory (MemGPT, Zep), and Verification (promptfoo, AgentBench).
• Numerous case studies from industry giants like Microsoft (Azure SRE Agent), Meta (Ranking Engineer Agent), and Google (ADK), showcasing production-grade harnesses.
• Focus on practical implementation patterns, including sandboxing (E2B, Daytona), observability (OpenLLMetry, Langfuse), and multi-agent orchestration (AutoGen, CrewAI).

Maintenance & Community

The project encourages contributions via CONTRIBUTING.md. Discussions and idea refinement are noted to have occurred within the linux.do community.

Licensing & Compatibility

The repository is dedicated to the public domain under CC0, allowing for unrestricted use, modification, and distribution. This facilitates broad adoption and integration into commercial and closed-source projects without licensing friction.

Limitations & Caveats

As a curated list, its primary limitation is that it points to external resources rather than providing direct tooling. The field is rapidly evolving, meaning some components may become obsolete as AI models advance. A key warning highlighted is the "co-evolution warning," where models trained with specific harnesses can become overfitted to those designs, implying that harness architecture choices have lasting consequences.