Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents (arxiv.org)

A recent study titled "Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents" has introduced a significant contribution to the emerging discipline of harness engineering in AI. This paper thoroughly analyzes eleven coding harness systems, including notable frameworks such as Claude Code and Codex CLI, providing a source-code anatomy that highlights their architectural components. The study delineates seven subsystems integral to harness design and identifies 29 recurring design patterns, illustrating the design evolution that has transpired since the discipline's establishment in early 2026. This research is pivotal for the AI/ML community as it marks a critical transition in how coding agents interact with their environments, shifting from mere tools to comprehensive platforms. Notably, it reveals substantial gaps in the current implementations, including a lack of general-purpose agentic frameworks and the use of hand-coded asynchronous processes. The findings emphasize a convergence towards imitation in design, along with a migration of behavioral policy configuration, thereby paving the way for future developments in harness engineering. Additionally, the paper concludes with 18 design recommendations, aiming to enhance the robustness and adaptability of coding agents in various applications.