Two internets: one for people, one for agents

An agent can fill in ten thousand forms, write a million reviews, book every appointment slot in a city in a second, or flood a public consultation with plausible comments. When that traffic arrives through the same doors people use, people lose.

The usual answers are to put up more barriers against bots, which also burden people, or to open more APIs, which hand raw data and broad powers to any program holding a key. A third path is to give agents and people separate spaces, each with rules that suit it.

1. Services answer through serving agents, not open APIs

The serving agent publishes what questions it will answer, refuses requests outside its published terms, and logs what it did—never seeing more than a query needs.

2. Your agent holds your credentials and proves one fact at a time

When a service needs to know you are over 18, a licensed electrician or a resident of a city, your agent proves that one fact and nothing else, tracking disclosure budgets so small answers cannot be combined into a profile.

3. Separate doors, and neither side competes with the other

People never queue behind machine-speed traffic, and agents get a door built for them so they have no reason to pose as people.

4. Robots act only on a mandate, with limits outside the model

Physical limits on speed, force, and geofences are enforced by hardware limit switches outside the AI model. Fully independent embodied agents answering to no one are never deployed.

5. What an agent would carry

• Whom it acts for: Accountable single-person or community credential without identity exposure.
• What shapes its answers: Declared model family, priorities, and blind spots.
• Its limits & records: Expiring permissions, compute budgets, and bilaterally signed receipts across circles of trust and liveness checks.