Public services that work for people and their agents.

Telling where a submission came from, and whether it states a view the user holds or one an agent generated. Nothing in a submission's text shows who wrote it, yet the agency still has to act on it.

Enforcing least agency when a user grants software standing authority to act. Identity systems prove who a person is, not the minimum scope they granted an agent or how to revoke it in flight.

Tracing every agent action back to a human decision-maker and halting systemic faults at runtime. Without pre-commit checkpoints and circuit breakers, a single flawed process repeats across a population before anyone sees it.

Helping a user judge how far to rely on a government AI agent, and giving the agency a basis to support that judgment. A user can wrongly refuse an agent that would help, or lean on one a serious case needs a human for.

Forms, queues, and deadlines rationed services for years by demanding effort no one designed in. AI agents strip that effort away, forcing a deliberate choice about which friction only excluded people and which metered need.

Reading the weight of opinion behind public submissions once agents file fluent, distinct-looking entries at scale. Volume and the look of variety can both be manufactured, so neither shows how widely a view is held.

Keeping every user able to reach a government outcome, from a consultation to a benefit, license, or appeal, as private agents come between them and government. Left alone, outcomes start to track who has the better agent.

Evaluating civic AI tools independently of the labs and vendors that build them. Building an authoritative-sounding agent is nearly free; verifying its reliability and backing it when it fails is not.

Governing sovereign exposure across the AI supply chain when a model runs, or can be legally compelled, outside the user's jurisdiction. It decides who can reach the data and whether the service can be relied on under pressure.