50 likes
Leaderboard ranks coding-agent harnesses for open-source models
What is harness-leaderboard?
The Harness Leaderboard runs a given model through nine coding-agent harnesses on Terminal-Bench 2.1 and DeepSWE v1.1 and records the resolve rate with confidence intervals. Each trial is executed in an isolated HF sandbox, and the full trajectories are stored publicly for anyone to inspect.
Who it is for
If you build or evaluate AI coding assistants, this lets you see which harness gives the best performance for your model without writing your own benchmark.
How to try it
Open https://huggingface.co/spaces/joelniklaus/harness-leaderboard