The open-source Python wrapper developed by Pawel Huryn achieved top success rates in a comprehensive series of tests on the Terminal Bench 2.1 coding benchmark. Huryn reported that his Mini Harness, which consists of roughly 1,400 lines of code, now outperforms other agents including OpenCode 2, mini-swe-agent, and Pi across all 89 tasks in the set. He plans to extend the benchmark trials to include the DeepSeek model and support for OpenAI endpoints.
The results build on a larger study by researcher Joel Niklaus, who analyzed the impact of different agent harnesses across 28,849 trials. Niklaus found that the mini-swe-agent had the highest overall solve rate at 51.4%, while the Pi harness proved most economical, costing $0.29 per solved task compared to $3.35 for opencode. While Huryn’s project previously only led in speed, the latest data shows it has moved from the middle of the pack to the top of the solve-rate leaderboard for specific configurations.
