Terminal-Bench 2.1

An Empirical Study of Harness Design for Coding Agents

A paper evaluates 176 matched settings across four models on SWE-Bench Verified and Terminal-Bench 2.1, isolating three harness components: planning, action space, and context management. It finds context management matters most when budgets tighten, staging rule-based elision before LLM summarization is most efficient, planning is an accuracy scaffold for weaker models but a cost saver for stronger ones, and bash-capable models can use a bash-only interface at lower cost.