I have this terminal bench results and you can a pretty dramatic difference between the top harness versus Claude Code. Do people think this is representative of what a good harness can do for tasks in terms of giving a major lift or are we seeing a lot of benchmaxing and these results over exaggerate the practical effect of a harness. I know we have something like this coming for OpenHand Index, but I need a slide to show the effects of a harness and this seemed the most contemporary example to use.

some of the top ones were even accused of cheating by baking the solutions to the problems into the docker images. i’m not sure if those are still on the leaderboard or not, but there is definitely some benchmaxxing going on. that being said, we should re-evaluate on terminalbench at some point