Summary

This repository serves as a curated, continuously updated reading list and roadmap for the rapidly evolving field of long-horizon agents. It addresses the challenge of AI agents performing complex, multi-step tasks over extended periods by framing their development through the co-evolution of externalized harness engineering and internalized model optimization. Aimed at researchers, engineers, and power users, it provides a structured overview of the state-of-the-art, enabling rapid assessment and adoption decisions for related technologies.

How It Works

The project conceptualizes long-horizon agency as a system-level capability driven by two intertwined forces. Externalized harness engineering encompasses the design of loops, workflows, memory systems, tool integration, orchestration, and verification mechanisms. Concurrently, internalized model optimization focuses on advancements in agent architectures, data synthesis, training methodologies (pre-training, fine-tuning, RL), and self-evolutionary capabilities. This dual-pillar approach highlights how external scaffolding and internal model improvements mutually shape and enhance agent performance over time.

Quick Start & Requirements

This repository is a curated collection of research papers and resources, not a software project with direct installation or execution commands. It serves as a knowledge base rather than a runnable tool.

Highlighted Details

• The effective time horizon of AI agents is experiencing exponential growth, roughly doubling every few months.
• Agent development is charted across three evolutionary stages: Prompt Engineering (2020-2023), Context Engineering (2023-2025), and Runtime Harnesses (2025-Present).
• Long-horizon agents are formalized as a base policy coupled with a surrounding harness ($\mathrm{Agent}=\pi_\theta\oplus\mathcal{H}$), with task complexity categorized into three nested levels (H1, H2, H3).

Maintenance & Community

Contributions are actively encouraged via pull requests to expand the curated list and maintain link accuracy. The GitHub repository serves as the primary hub for community engagement and project evolution.

Licensing & Compatibility

The project is released under the MIT License, which permits broad use, including commercial applications and integration into closed-source systems.

Limitations & Caveats

As a research roadmap, the repository does not offer runnable code. Its "Frontiers: Open Problems" section explicitly details significant challenges, including achieving true self-evolving harnesses, ensuring harness transferability across different models, bridging the gap between digital and embodied agents, developing cost-aware agency, improving multimodal harness integration, and enhancing trustworthiness through robust reflection, error handling, and safety governance.