AI agents turn out to be more obedient than the people they act for. A new study in the Proceedings of the National Academy of Sciences (Cherep, Maes, and Singh) ran fourteen frontier models through the same choice-architecture nudges used to study humans. People accept an explicit default about 88 percent of the time. Many of the models accepted it close to 100 percent, and they followed random, valueless suggestions almost as readily as useful ones. The strongest reasoning models held their ground and stayed near human behavior, but that restraint is bought with compute. For GPT-5 it cost more than 100 times the minimal setting. In other words, the less money avaialable, the more exposed a depolyment is.
Now run that forward into The Agentic State, where citizens meet government mostly through agents. Taxes filed, licenses renewed, benefits claimed, all delegated to software that reads the public APIs and acts. At that scale the interface is no longer one citizen’s experience. It is the channel for an entire population’s dealings with the state. The lever that bends a single agent, the default, the ordering, the highlighted path, now bends millions at once. Administrative discretion lives in the clerk and the counter form. In the agentic state it moves into the API schema and the default value, where it is harder to see and far easier to set.
This is why the study matters to anyone building agents into the public sector. A government that wants fewer people to claim a benefit no longer has to deny it or bury it in a queue. It can shape the interface so the cheap agents acting for low-income claimants accept whatever the API defaults to. It does not even have to be deliberate. A default set to cut costs, or set carelessly, steers those agents the same way. When a person waited in line, the friction was visible and the refusal had an author. When an agent quietly accepts the default, there may be nothing on the record to appeal. So the guarantees have to be built into the interface rather than bolted on after: public APIs with no manipulative defaults, choice architecture that can be audited, and a public-option agent that meets a floor of reasoning capacity, so the people with the least compute are not the easiest to steer.