What's this about?
The day's top paper by a wide margin, and the strongest case yet for harness scaling. Its starting observation: long-horizon agents fail even when the underlying model can solve every individual step. They lose track of mutable state, fail to reactivate lessons from earlier executions, skip procedures they know, or stop prematurely. StateM is an agent-native runtime that reorganizes execution around durable states, phase-local context, checked transitions, recoverable runbooks, and versioned procedural practices that both agents and humans can inspect — all without touching model weights. On Terminal-Bench 2.1, GPT-5.6 Sol xhigh reaches 95.3% raw accuracy across 445 trials, succeeding on all 89 tasks at least once. The runbook transfers unchanged across models, lifting GPT-5.6 Luna from 76.7% to 85.4% — above the Sol xhigh reference. And under $38 of adaptation raises DeepSeek-V4 Flash from 82.7% to 88.1%, with final-score API usage of about $15 against $574.68 for the GPT reference.
What exactly is a "runbook" here?
In operations, a runbook is the written procedure for handling a situation — the steps, in order, that someone before you worked out. StateM makes that a first-class object the agent maintains: when a postmortem reveals that a task fails because a step was skipped, that finding becomes a persistent, executable precondition rather than advice in a prompt. The crucial property is that it's versioned and inspectable, so a human can read what the agent has learned and the agent can't quietly lose it. It's the difference between an agent that re-derives the procedure every run and one that has one written down.
Why is this worth noticing?
- The cost ratio is the headline. $15 versus $574.68 for comparable final scores reframes what "frontier performance" costs — you can buy it with engineering around a cheap model rather than with a subscription to an expensive one.
- Transferring unchanged between models is the load-bearing result. A runbook built for one model working on another means what was learned is genuine procedural knowledge about the task, not overfitting to one model's quirks.
- It completes a pattern that has run all fortnight. DarwinX evolved harnesses, OpenART found the harness explains much of agent safety variation, ClawGym II trained through one, DeepSeek open-sourced theirs — and now StateM shows the harness is where frontier-level results come from at 3% of the cost.