AI agents and the trust budget: why mobile networks matter

An autonomous agent walks into a website. The page loads. The agent fills the form. The page says “we noticed unusual activity from your network.”

That is not a model problem. That is a trust problem.

When teams ship their first browser-using agent, they spend most of the engineering effort on reasoning, planning, and tool use. The first time the agent is asked to do something a real user would do, the model is fine. The route in is what trips.

The trust budget

Every interaction with a destination spends a small amount of trust. The starting balance is set by the address you are connecting from, the network reputation of that address, and the headers and timing of your request. The balance ticks down as the destination sees behaviour that diverges from what its average user does.

When the balance hits zero, the destination starts asking questions. It shows a verification step, throttles, or simply returns empty content. Most agents are not designed to recover from that gracefully.

You do not negotiate with the trust budget. You start with a bigger balance.

Why this favours mobile routes

The trust balance for a carrier-grade mobile address is large. Mobile networks share addresses between many real users, so a destination knows that any given address has a long history of normal browsing behaviour behind it. The same destination treats a fresh datacentre address with suspicion immediately.

That is why our recommendation for agentic browsing is Mobile when the destination is one of the harder consumer sites: search, marketplaces, social platforms, login-gated dashboards. The math is simple. A mobile route gives the agent the room to behave like a user before the destination starts asking questions.

A model that can plan ten steps is not useful if the destination disconnects after step two.

When ISP is the better default

Not every agent needs the full trust balance of a mobile carrier. A research agent reading public news pages, a translation agent pulling product descriptions, a comparison agent walking a catalogue — these tasks rarely trigger the trust spend at all.

For those, Platinum is the sensible default. The route is clean, the city and ASN are correct for the user the agent is impersonating, and the costs are predictable per gigabyte. You spend less, and you do not need the resilience of a mobile route to complete the job.

The rule of thumb: if the agent is reading, ISP is usually enough. If the agent is acting on behalf of a person and the destination expects a person, mobile is where you start.

Holding the route

Agents often need to keep the same route across many steps. A login is on one address. The dashboard load that follows must be on the same address or the destination invalidates the session. The same applies to shopping cart flows, multi-form submissions, anything stateful.

We expose sticky sessions on every plan for exactly this reason. The agent appends a session token to the credentials and keeps it for as long as the workflow needs. When the workflow ends, the token is dropped and the next workflow starts fresh.

This is the second half of trust budget management. Spend it on one identity per workflow, not many.

What to measure

Two numbers help more than any others.

Completed-workflow rate: how often the agent finishes the multi-step task, not just the first request. This is the real failure signal. A 95 percent success rate on the first request masks a 40 percent completion rate on the workflow that the request was supposed to start.

Trust-recovered rate: when a workflow stalls, how often does retrying on a new session on the same network type recover it. If that recovery rate is high, the network choice is correct and the model needs a better retry plan. If it is low, the network choice is wrong for the destination.

Run both for a week. The right answer usually picks itself.

A starting question

If you are deciding where to start, ask one question first: does the destination expect a real person, or a researcher? The answer points at mobile or ISP, and the rest follows.

When in doubt, start on the higher-trust network. You can move down to ISP once you have evidence that the destination tolerates it. Moving up the other way, after a destination has already flagged your address space, is much harder than just starting there.