Researchers conducted an empirical study on harness design for autonomous coding agents, evaluating how individual components affect software-engineering performance. Using a lightweight harness with fixed execution loops, they tested 176 matched settings across four models on SWE-Bench Verified and Terminal-Bench 2.1, varying planning, action space, and context management components.

Key findings include: context management becomes increasingly valuable as context-window budgets tighten, primarily by preventing context-overflow failures. Staging rule-based elision before LLM-based summarization provides the strongest efficiency, while making elided content recoverable adds complexity without accuracy gains. Planning shifts from an accuracy scaffold for weaker models to a cost saver for stronger models with minimal accuracy change. Predefined tools improve performance for models with weaker bash proficiency, while bash-capable models achieve substantially lower costs with a bash-only interface, especially on command-line-centric tasks.

Trajectory-level analysis reveals context management extends execution trajectories without altering agent behavior, planning changes where trajectories stop, and action space changes code-writing granularity. These findings inform model- and budget-aware harness design and provide a modular framework for evaluating future harness components. The study was submitted by Run-Ze Fan on September 17, 2026.