Study Dissects Which Coding-Agent Harness Components Actually Help
The paper runs a controlled, component-level study of coding-agent harnesses rather than treating them as black boxes, varying planning, action space, and context management independently across 176 settings, four models, and two benchmarks (SWE-Bench Verified, Terminal-Bench 2.1). Key findings: context management mainly helps by preventing context-overflow failures and matters more as context budgets shrink; rule-based elision before LLM summarization is the most efficient context strategy; planning helps weaker models' accuracy but mainly saves cost for stronger ones; and predefined tools aid models with weak bash skills, while bash-capable models do fine (and cheaper) with just a bash interface. Trajectory analysis ties these effects to concrete behavioral mechanisms—trajectory length, stopping points, and code-edit granularity.