Agent-native payments

The framing that makes this more than a crypto pitch: an agent stops at a paywall or a metered API not because it cannot decide but because it cannot pay. That makes the natural interface a protocol status code the agent handles inline — a price returned with the refusal — rather than a checkout flow built for a human with a card.

Per-wallet programmatic rules, absolute rather than probabilistic finality, sub-cent settlement, and selective rather than total transparency. The demonstrated economics are the argument: a cent for a registry-grade domain check, six cents for a specialist image model, tens of cents for a real phone call — amounts that card rails cannot process at all, since percentage fees plus per-transaction minimums destroy cent-scale work.

Spend limits bound the blast radius of a bad purchase, not the judgment behind it, and nobody has formal verification of what an agent decides to buy. That caveat comes from the party selling the rails, which is the most useful sentence in the cluster.

Also: the x402 protocol; EIP-3009; HTTP 402; and the GENIUS Act’s rulemaking.

Both talks are from the same company, so the deployment numbers are a vendor’s own and there is no independent corroboration in this scan. That is why the concept stays emerging despite real numbers.

The “human one step away” assumption these rails are designed to remove is precisely the one Assay keeps on purpose: human-gated merges, human-only closure on decision issues, and no ambient spending authority anywhere in the desk verbs. What does transfer is the shape of the control rather than the payment itself — per-wallet programmatic rules are the same idea as narrow per-role verbs bound to a roster identity: capability with logic attached, rather than ambient authority. Assay arrived at that independently.

The rails would matter only if a desk ever needed to procure something at run time — paid eval endpoints, premium data, or per-call specialist models. Nothing in the current stream set requires it. If that changes, the vendor’s own caveat is the one to carry forward: a spend limit is a blast-radius control, not a judgment control, so a purchasing desk would need the purchase decision gated the way a merge is, not merely capped.

• Whether a second, independent organization reports agent-paid volume — the single-vendor evidence base is the main reason this stays emerging.
• Whether any HTTP-402-based scheme is adopted by an API provider outside the sponsoring marketplace, which is what would make it a protocol rather than a product.
• Whether anyone attempts to gate the decision to purchase rather than the amount — that is the open problem the vendor names and nobody in this scan addresses.