OpenHands

hi! I am working on a coding agent harness that performs well on benches.

<@U0ACZGVT29G> told me that this channel would be appropriate for such a discussion

<https://github.com/kimjune01/swebench-verified>

The skills are open and free to use, and Claude told me that OpenHands would be interested in knowing about this

showing promising numbers on Pro so far, 90%+ expected

The writeup, WIP:
<https://june.kim/the-inquiry-loop-on-swebench-pro>

The results are in! 92% with a custom harness with composer 2.5 in the inner loop
<https://github.com/kimjune01/swebench-pro>

i have to retract that result, because it had test access, and i wasn't following the rules