MacAgentBench: Benchmarking AI Agents on Real-World macOS Desktop

MacAgentBench is a new benchmark for evaluating computer use agents (CUAs) on macOS, featuring 676 tasks across 25 applications, with a focus on both GUI and CLI interactions. It employs deterministic rule-based evaluation and fine-grained multi-checkpoint scoring, revealing that Claude Opus 4.6 on OpenClaw achieves a 73.7% Pass@1 score, primarily due to its skill library. This benchmark is significant for practitioners as it provides a more comprehensive assessment of agent performance in real-world scenarios, highlighting the importance of framework capabilities in long-horizon, multi-application tasks.