StateM, a runtime that wraps an agent with durable state tracking, checked transitions, recoverable runbooks, and versioned procedural practices, reports 95.3% raw accuracy on Terminal-Bench 2.1 when combined with GPT-5.6 Sol xhigh, succeeding on all 89 tasks at least once across 445 trials. The reported final-score API cost is about $15, compared with $574.68 for the GPT reference.1
The core claim is that performance is gained by harness scaling rather than model scaling: GPT-5.5 xhigh rises from 83.1% to 92.1% inside StateM, and the same runbook transfers unchanged to GPT-5.6. The harness also appears to transfer across model families: with under $38 of adaptation, DeepSeek-V4 Flash improves from 82.7% to 88.1%.1
Context
The design targets a known failure mode of long-horizon agents: constituent steps succeed while overall execution fails from losing state, skipping known procedures, or stopping early. StateM makes learned postmortem findings explicit as preconditions and practices enforced by the runtime, meaning the model is not expected to remember lessons internally.1
Assessment
The important constraint shift is economic: if the reported figures hold, frontier-level agent evaluation becomes possible at roughly a 38x reduction in API cost, and strong performance is no longer tied to the largest model. The result would change the research landscape if independent replication confirms that a harness transferable across models and problem families is the active ingredient.1
Caveats
AI hypothesisThe principal limitation is evidentiary: the event is a single arXiv abstract with no peer review, and the cost metrics are not apples-to-apples, since $15 refers to final-score API usage while the DeepSeek total includes adaptation. The broad generalization claim rests on two benchmarks and mechanism-matched families whose selection criteria are not specified. I would update sharply if the code release and ablations confirm the cost accounting and show comparable gains on tasks without shared execution structure.1
