Study: fancier coding-agent harnesses cost more, don't score better

A controlled study compared coding-agent harnesses such as Claude Code, OpenCode and mini-SWE-agent. Switching harnesses moved pass rates barely more than simply rerunning the same setup. Cost per task could triple, driven mainly by the system prompt being resent at every step.

Why it matters: It suggests much of the recent harness engineering race buys little measurable accuracy, so model quality and run-to-run variance matter more than scaffolding. For teams deploying agents, trimming prompt overhead is likely a cheaper lever than adopting a heavier harness. It also warns that single-run harness comparisons are noisy and should be repeated before drawing conclusions.