Mouse tops FrontierHarness at 24/30 with Kimi K3 using completion loops (mouse.dev)

In a recent benchmark called FrontierHarness, the Mouse coding agent achieved an impressive score by passing 24 out of 30 tasks with the Kimi K3 model, which is the highest pass rate recorded to date. This benchmark consists of two sets of tasks: Terminal-Bench jobs, which involve command-line operations, and DeepSWE tasks that deal with real-world source code challenges. The pioneering aspect of Mouse lies in its unique completion loop mechanism, built on the OpenCode engine, which allows it to iteratively verify task outputs until they meet defined criteria, showcasing significant advantages for complex, long-horizon jobs. The significance of Mouse's performance for the AI/ML community is underscored by its ability to outperform other agents, especially on DeepSWE tasks, where it scored 6 out of 9 compared to 0 for OpenCode. This differentiates it in a landscape where the precision of coding agents is critical for automating software development and bug fixing. While its cost per pass and time efficiency are areas for improvement, the innovative approach implemented in Mouse could pave the way for future advancements in coding automation, demonstrating a promising direction for high-performance AI tools in software engineering.