An Empirical Study of Harness Design for Coding Agents
A controlled study treats coding-agent harnesses as composable systems: 176 matched settings vary planning, action space, and context management across four models on SWE-Bench Verified and Terminal-Bench 2.1. The strongest result is practical: staged rule-based context elision before LLM summarization gives the best efficiency, while recoverable elision adds machinery without improving accuracy. Planning mainly scaffolds weaker models but saves cost for stronger ones; bash-capable models can often use a bash-only interface more cheaply than predefined tools.