Research · Sunday 25 October 2026 · story 13
Study finds coding agent harnesses raise cost more than results
THE DECODER reports a study found Claude Code and mini-SWE-agent stayed statistically close when the model was fixed, while OpenCode lagged. It says harness swaps often looked like rerun noise, but per-task costs could be up to three times higher.
Why it matters. For coding agents, this suggests harness changes can raise token spend more than pass rate, so reruns matter before treating small leaderboard gaps as meaningful.
Read the original at the-decoder.com