Frontier models can manage grants, review compliance, underwrite risk and run laboratory protocols. Adoption has stalled anyway, at the line between assisting a person and working alone.
We argue that two separate failures hold agents back: losing a rule from the context, and keeping the rule but acting against it. They need different remedies, and repeating the rule cures only the first.
What we found
- The barrier is trust. Longer prompts, bigger context windows and reasoning models each make failures rarer; none makes them impossible.
- On capable models the second failure largely disappears: a rule the model can see holds, even under pressure. What remains are rules that have left the context and rules nobody wrote down.
- Unwritten rules leave the widest gap. 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.