Executive summary
Three in four enterprises that put AI agents in front of customers have pulled them back, according to one industry survey. The Durability Frontier asks why that happens when frontier models can out-reason most human operators, and traces it to commitments: agents lose binding rules over the length of real work. An agent economy built on capable models needs agents that keep their commitments.
Findings
- The production systems the paper surveys all treat agent memory as storage. Storage helps a model state a rule correctly; acting on it takes enforcement, which none of them provides.
- Forgetting happens all at once. Within the 12,000 tokens tested, the model read a rule perfectly however far back it sat, and a rule that had scrolled out weighed no more than one never given.
- Searching for similar text added nothing. On a regulated task with four kinds of constraint (debarment, budget caps, subsidiary structure and eligibility waivers), an agent searching for similar text scored 0.429, the same as one with no help, because the rule that governs a request often shares few of its words.
- Nenya held that task at 1.000. The strongest alternative tested, LangGraph, reached 0.886 at short horizons and 0.639 over long ones.
Nenya keeps the rules you set in force for the whole job, including the stretches where the original instruction has scrolled out of view.