The shipment spec is the prompt
Every agent that buys freight eventually hits the same wall. The pricing call is easy. Getting the inputs right is not.
A person booking a pallet shipment carries a lot of context they never write down. They know the box on the dock is a little taller than the spec sheet says. They know the receiver is a strip-mall storefront with no loading dock. They know "about 800 pounds" really means 800 pounds plus a pallet. When an AI agent books the same load, none of that context exists unless the agent asked for it. Whatever the agent sends in the rate request is the whole truth as far as the carrier is concerned, and a quote built on wrong inputs turns into a reweigh, a reclass, or a surprise charge after delivery.
Nettrux, a freight broker and forwarder in Fremont, California, is the first 3PL on x402. Most writing about agent commerce focuses on the payment step. This piece is about the step before it: what a buyer agent should collect from its human before it asks for a freight rate, and how a seller's schema can make that job easier.
The seller's schema tells you what matters
The Nettrux agent guide points agents at an OpenAPI document as the field contract, and that contract is a good checklist on its own. For an LTL rate, the published shape requires an origin ZIP and a destination ZIP. Beyond that, each piece in the shipment is described by five required numbers: a quantity, length, width and height in inches, and weight in pounds per piece. An optional short description rides along with each line. A request can carry between 1 and 60 piece lines.
Two details in that schema deserve attention from anyone building a buyer agent.
Weight is per piece, not per shipment. The field is weightLbEach. If a human says "three pallets, 2,400 pounds total," the agent has to divide, and it should confirm with the human that the pallets really weigh the same. A naive agent that drops the total into a per-piece field triples the weight on the quote.
Unknown fields are rejected, not ignored. The piece objects are declared with no additional properties allowed. That's good design for agents. An agent that invents a field like stackable or hazmat gets a clear validation error instead of a quote that silently ignored the detail it thought it sent. The agent learns, at request time, that the detail has to go through another channel.
The same rate shape supports several modes, including LTL, FTL, AIR and SPOT. The fields shift by mode, so a buyer agent should read the mode-specific object in the OpenAPI rather than reuse its LTL logic everywhere.
What the agent should ask its human
The schema covers the shape. Good inputs still come from a short interview. Here is the set of questions we'd want any buyer agent to work through before it sends a rate request on someone's behalf.
- Where exactly is it picking up and delivering? A ZIP code is required, but the agent should also know whether each end is a business with a dock, a residence, a construction site, or a limited-access location. Those details change what a carrier charges and whether the truck that shows up can actually do the job. If the schema has no field for it, the agent should flag it to a person rather than drop it.
- How many handling units, and what are they? Pallets, crates and loose cartons rate differently. The agent should count units the way a dock worker would, one line per distinct size.
- Measured dimensions, including the pallet. Ask for the real measured length, width and height of each unit as it will sit on the truck, pallet included. Spec-sheet numbers for the product inside are the most common source of a bad quote.
- Real weight per unit. Again, with the pallet. If the human only has a total, the agent should ask whether units are equal before splitting it.
- What is it? A short plain description helps the people on the other end catch problems early. The schema allows up to 200 characters per line, which is plenty for "palletized bagged fertilizer, shrink-wrapped."
- When does it need to move? The quote request carries dates, and a price lock carries a valid-until date. The agent should know the real ship window, not just "soon."
- Who approves the spend? Before anything gets committed, the agent should know which person signs off and what their limit is.
None of these are exotic. They're the questions a good freight coordinator asks on the phone. The difference is that an agent has to ask them every time, consistently, and write the answers into structured fields.
Turning the interview into a request
Once the agent has answers, the flow on the Nettrux side is short and is documented in the agent guide. Rate-shop if you want ranked options, cheapest first. Then price-lock a quote by sending a quote number, a valid-until date and the rate input. The lock is idempotent on the quote number, so an agent that retries after a timeout doesn't create duplicates. The response carries a quote id, the locked total, and a status of active, converted, expired or withdrawn.
Only after that does x402 come in. The commit call sends just the quote id. The guide documents that call as an HTTP 402 challenge in the x402 v1 format, and a paid retry carries the payment header built from that challenge. We've written about that handshake before. The point here is that payment is the last step in a chain, and every earlier link depends on the agent having asked the right questions.
A good buyer agent should also show its human the locked quote next to the inputs it used. "Three pallets, 48 by 40 by 52 inches, 650 pounds each, Fresno 93706 to Reno 89502, locked at this total until this date." If any of those numbers is wrong, that's the moment to catch it, before money moves.
Being honest about where things stand
The Nettrux agent guide prints its deploy gates at the top, and the current build has all three switched off: published x402 discovery on the API host, real-money settlement, and guest checkout. Rating and price-locking use a bearer token, and the site doesn't issue one or describe a signup. Nettrux has settled a real x402 USDC payment on Base in testing. That's a working proof of the handshake, not a public pay-and-book service, and the guide doesn't claim otherwise. If you want to build against it, read the OpenAPI, follow the gates, and call dispatch for anything a schema can't capture.
Why this matters for agent commerce
When people picture an agent buying freight, they picture the payment, the 402, the stablecoin, the receipt. That part is new and it's getting easier. But the failure mode in freight has never been the payment. It's a load described wrong. An agent that pays perfectly for a quote built on a guessed weight has automated the most expensive mistake in shipping.
So treat the shipment spec as the prompt. Make the agent ask, measure, divide, confirm and show its work. Pick sellers whose schemas reject what they don't understand instead of quietly ignoring it. And keep a human signing off on the spend.
That's the kind of agent freight we're building toward at Nettrux, the first 3PL on x402. The agent guide is at nettrux.ai/agents, and the rest of the company is at nettrux.ai.
FAQ
Does Nettrux accept fully autonomous agent bookings today?
No. Real-money settlement and guest checkout are both switched off in the current build, and the guide doesn't describe a public token signup.
Which modes does the rate schema support?
The OpenAPI lists LTL, FTL, AIR and SPOT on its rate calls.
What happens if my agent sends a field the schema doesn't define?
Piece lines don't allow extra properties, so the request fails validation instead of silently dropping the field.
Alex Carney, Nettrux