StateM Runtime Achieves 95.3% Accuracy on Terminal-Bench 2.1

A recent study published on arXiv (2608.15089) presents StateM, a runtime specifically designed for agents that enhances the performance of long-horizon AI agents without modifying their model weights. This system structures execution around persistent states, context relevant to specific phases, verified transitions, recoverable runbooks, and version-controlled procedures. In tests with Terminal-Bench 2.1, StateM improved GPT-5.5 xhigh accuracy to 92.1%, up from a baseline of 83.1%, and achieved 91.9% with GPT-5.6 Sol Ultra. The unchanged runbook also applies to GPT-5.6. With GPT-5.6 Sol xhigh, StateM reached 95.3% accuracy over 445 trials, completing all 89 tasks at least once. Additionally, the frozen profile enhanced GPT-5.6 Luna from 76.7% to 85.4%, exceeding the 84.9% Sol xhigh benchmark. The paper notes that long-horizon agents frequently struggle despite their models' capabilities, often losing track of mutable states or failing to leverage past lessons. StateM counters these challenges by offering a well-structured runtime that is accessible to both agents and users. The research emphasizes 'harness scaling' as a viable alternative to model scaling, indicating that advancements in execution infrastructure can lead to notable performance improvements.