Most long-horizon agent failures I’ve traced aren’t reasoning failures. The model could solve every step in isolation. What breaks is everything around the step: the agent forgets a mutable value it set three tool calls ago, re-runs a procedure it already finished, or stops before the task is actually done. StateM makes exactly that its thesis — scale the harness, not the weights.
The runtime organizes execution around durable states, phase-local context, checked transitions, and recoverable runbooks, plus versioned procedural practices that both the agent and a human can inspect. That last piece is the tell. Treating the execution trace as a first-class, inspectable artifact is precisely what you need when an agent run goes sideways in production and someone has to reconstruct why after the fact. It reports 95.3% raw accuracy on Terminal-Bench 2.1 for roughly $15 a run, but the number matters less than the mechanism that produced it.
This lands where I keep landing on multi-agent orchestration: the leverage is in state management and checked transitions, not in another point of raw model capability. If your agent loses the thread on step 12 of 20, a smarter model buys you step 13 — a better harness buys you all of them. The HF paper page has the full breakdown. The question I’d want answered before adopting any of this: how much of that 95.3% survives when the runbook itself is wrong and the agent has to recover from a bad procedure, not just a bad step?