StateM: Reaching 95.3% Raw Accuracy, or a \$15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling

Long-horizon agents can fail even when their underlying models can solve the constituent steps. They may lose track of mutable state, fail to reactivate lessons from earlier executions, skip known procedures, or stop prematurely. We bet on harness scaling to improve the execution system around an agent without changing its model weights. We introduce StateM, an agent-native runtime that organizes execution around durable states, phase-local context, checked transitions, recoverable runbooks, and versioned procedural practices that agents and users can inspect together. On Terminal-Bench 2.1, GPT-5.5 xhigh with StateM reaches 92.1% accuracy, compared with an 83.1% reference and surpassing GPT-5.6 Sol Ultra (91.9%). The same runbook transfers to GPT-5.6 without modification. With GPT-5.6 Sol xhigh, the system records 95.3% raw accuracy across 445 public-submission trials and succeeds at least once on each of the 89 tasks. The frozen profile also raises GPT-5.6 Luna from 76.7% to 85.4%, numerically above the 84.9% Sol xhigh reference. Next, starting from the same runtime, runbook structure, and golden rules, it takes less than $38 of adaptation cost to raise DeepSeek-V4 Flash from 82.7% to 88.1% on the full benchmark under standard timeouts.

Introduction. Long-horizon agents often fail in a revealing way: the underlying model appears capable of solving each local step, yet the complete run still fails. The agent deviates from its plan, loses track of mutable task state, skips a necessary check, repeats an unproductive action, or stops before the requested deliverable is verifiably The dominant response to such failures is to improve the model: scale pretraining, add post-training data, increase test-time reasoning, or introduce additional agents. We study an orthogonal question: We call the systematic improvement of this surrounding control layer harness scaling. Harness scaling does not replace model scaling. It asks whether more of a model’s existing capability can be converted into completed, reliable work by improving the runtime around it. This leads to three progressively stronger empirical tests: first, can a better harness improve a fixed model without changing its weights? Second, can a harness developed with one model transfer to a newer model without retuning?

Discussion / Conclusion. Harness scaling requires a runtime that is both agent-native and enforceable. The experiments above highlight a practical requirement that is easy to miss when viewing a harness only as a fixed workflow: the runtime itself must be cheap to modify. Conventional state-machine and graph-based agent runtimes provide explicit orchestration and strong control, but typically externalize substantial reasoning into manually constructed nodes, edges, and handlers. As the control policy evolves, modifying this structure becomes an engineering task of its own, making rapid failure-driven iteration increasingly heavy.