Executive summary
Frontier models are capable enough to manage grants, review compliance, underwrite risk and run laboratory protocols. Adoption has stalled anyway, at the boundary between assistance and autonomy. This paper argues that two separate failures hold agents back: losing a rule from context, and keeping the rule but acting against it. They need different remedies, and repeating the rule cures only the first. Autonomous work needs both handled.
Findings
- The barrier is trust. Longer prompts, bigger context windows and reasoning models each make failures rarer, and none makes them impossible.
- On capable models the second failure largely disappears: a rule they can see holds, even under pressure. What remains are rules that have left the context and rules nobody wrote down.
- Unwritten rules are where the gap is widest. They exist only in the pattern of past cases, so there is nothing to restore, and an unaided agent repeats the same violation each time the case comes round.
Nenya handles the failures capable models leave open, from rules that have scrolled out of context to rules nobody wrote down.