Coding-agent harness choice reportedly affects cost more than accuracy
Arena says UC Berkeley PhD candidate Melissa Pan explored the “harness tax”: how the system around an AI model affects its cost and performance.
TLDR
Arena frames its discussion around Claude Code, Codex and Pi. It says Pan reported three findings about coding-agent harnesses, including that harness choice impacts cost more than accuracy. Arena links to a video for the other two findings.
Combined views
17.6K
3 Sources, first seen 2h ago
likes
