An Empirical Study of Harness Design for Coding Agents
Summary
This paper empirically evaluates harness design for autonomous coding agents, focusing on planning, action space, and context management. It finds that context management is especially beneficial under tight context windows, that elision-based context-management strategies improve efficiency, and that planning and tool usage interact with model capability to affect cost and accuracy. The work provides a modular framework for comparing harness components to guide budget- and model-aware automation design.