Hugging Face researchers have introduced StateM, an agent-native runtime that enhances long-horizon agent execution through durable states, recoverable runbooks, and enforceable procedural controls—all without altering model weights. The system addresses common failure modes where agents lose track of state, fail to apply lessons from earlier runs, skip procedures, or stop prematurely.
On Terminal-Bench 2.1, StateM raises GPT-5.5 xhigh to 92.1% accuracy from an 83.1% reference, and GPT-5.6 Sol Ultra to 91.9%. With GPT-5.6 Sol xhigh, the system achieves 95.3% raw accuracy across 445 trials and succeeds on all 89 tasks at least once. The runbook transfers unchanged to GPT-5.6, and the frozen profile lifts GPT-5.6 Luna from 76.7% to 85.4%, surpassing the 84.9% Sol xhigh reference.
StateM also demonstrates cross-model generalization: with less than $38 of adaptation, DeepSeek-V4 Flash improves from 82.7% to 88.1% under standard timeouts and to 89.1% on an 88-task common core. Extending only the remaining latency-sensitive task matches the reported 88.8% GPT-5.6 Sol max result. Final-score API usage is about $15 versus $574.68 for the GPT reference; total DeepSeek expenditure is $52.22.
On BusinessBench, family-specific runbooks built on development sets yield held-out gains of 0.55 macro and 1.34 micro points; two mechanism-matched families improve by 10.04 points. The researchers argue that concrete rules generalize when tasks share execution structure, while the control methodology applies broadly. Code is available at github.com/henryqin1997/statem.