Three in four enterprises that put AI agents in front of customers have pulled them back, according to one industry survey. Frontier models can out-reason most human operators, so we asked why.
We trace it to commitments. Over the length of real work, agents lose the binding rules they were given, and an economy built on capable agents needs agents that keep them.
What we found
- The production systems we surveyed all treat an agent’s memory as storage. Storage helps a model state a rule correctly; acting on the rule takes enforcement, and none of them provides it.
- Forgetting happens all at once. Within the 12,000 tokens we tested, the model read a rule perfectly however far back it sat. Once the rule scrolled out, it weighed no more than one that was never given.
- Searching for similar text added nothing. On a regulated task with four kinds of constraint, an agent with similarity search 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 the task at 1.000. The strongest alternative we tested reached 0.886 over short horizons and 0.639 over long ones.