StateM was submitted to arXiv on August 15, 2026 (paper 2608.15089) by Ziheng Qin, Yaxin Lu, Zhangyang Atlas Wang and Kai Wang. It argues that a large share of long-horizon agent failure is not a model deficiency at all but a harness deficiency, and that the remedy is what the authors call harness scaling: improving the execution environment around a model without modifying its weights.
StateM itself is a runtime built from five ingredients: durable state that survives across steps, phase-local context so the agent is not carrying its entire history into every decision, checked transitions between phases, recoverable runbooks that let a failed step be resumed rather than restarted, and versioned procedures. The reported results on Terminal-Bench 2.1 are that GPT-5.6 Sol at extra-high reasoning reaches 95.3 percent raw accuracy across 445 trials, and that DeepSeek-V4 Flash reaches 88.1 percent with minimal adaptation cost, put at 38 dollars. The cost figure the authors lead with is the final API usage of about 15 dollars versus 574.68 dollars for the GPT reference run. On BusinessBench, they report held-out improvements ranging from 0.55 to 10.04 points depending on the task family.
Two caveats belong next to those numbers. First, they are the authors’ own, on a benchmark whose difficulty is calibrated by its maintainers rather than by an independent adversary, and a near-95 percent score on any agent benchmark should raise the question of whether the benchmark is now saturated. Second, the BusinessBench spread of 0.55 to 10.04 points is wide enough to suggest the technique helps a lot on some task shapes and barely at all on others.
For a technical leader, the strategic point survives the caveats. If a scaffold change can move a fixed model from a reference score to near ceiling while cutting cost by an order of magnitude, then agent performance in production is substantially an engineering variable rather than a model-selection variable. That reframes a common buying decision: the money currently spent chasing the next frontier model may be better spent on state management, recovery and context discipline in the harness you already run. It also means benchmark leaderboards that report model names without reporting harnesses are describing only part of the system being measured.