The words, pinned down

An open protocol for agent-to-agent communication: how one agent discovers another, describes its capabilities, and exchanges tasks with it — regardless of vendor or framework. Where MCP connects an agent to tools and data, A2A connects agents to each other. Payments enter the picture when the task one agent hands another costs money.

An open specification (OpenAI and Stripe, 2025) for completing purchases inside AI chat surfaces: the merchant remains merchant of record, the agent platform carries the checkout conversation, and payment credentials travel as scoped, single-use tokens. The commercial logic is that the chat surface becomes a storefront; the governance logic — delegated tokens with narrow scope — rhymes with the mandate pattern everywhere else in this lexicon.

A payment initiated by software acting on a person’s (or organization’s) behalf, under delegated authority, rather than by the person at a checkout. The defining problem is not moving the money — existing rails do that — but proving, pricing, and policing the delegation: who authorized this agent, for how much, on what, and who owns the error when it buys the wrong thing.

The consumer-legible model for delegated money: a cap (how much), a scope (on what, from whom), an expiry (until when), and a revoke (one gesture, immediate). Contrast with giving an agent a wallet — a custodial balance that answers none of those questions and concentrates custody and float risk. “Allowances, not wallets” is this site’s first principle (cut-0101).

An open protocol, announced in 2025 by Google with a broad coalition of payments and technology partners, for authorizing and accounting for agent-initiated payments. Its central object is the signed mandate — a verifiable record of user intent with scope and constraints — designed to travel alongside transactions across otherwise different rails. Payment-method-agnostic by design; extends the A2A/MCP family of agent protocols.

A signed statement by which one party vouches for a fact about another: that an agent belongs to a registered operator, that a mandate was signed by a verified human, that a piece of content came from a named publisher. Attestations are how trust crosses organizational boundaries without shared databases. This site practices what it defines — every Cut’s machine twin carries an Ed25519 signature verifiable against /.well-known/agenticpayments.json.

The card system’s consumer-protection weapon: a cardholder disputes a transaction and the money is clawed back through the network’s adjudication process, with liability assigned by rulebook. Merchants experience it as a tax; cardholders experience it as the reason they’ll hand a 16-digit number to a stranger’s website. For agentic rails the chargeback is the uncomfortable benchmark: the dispute layer most of them haven’t built yet (cut-0213, forthcoming).

A stored-value balance spendable only within the issuer’s own ecosystem — a platform credit, a game currency, a prepaid coffee card. Economically: the platform holds float, earns lock-in, and escapes some regulation open wallets attract, until scale attracts it anyway (see PPI). Structurally identical to the prepaid API-credit balances of AI platforms — which is why the PPI-India playbook keeps predicting their future (cut-0142, forthcoming).

A payment topology where the payer’s side initiates the transfer (“send ₹340 to them”), versus debit-pull, where the payee presents credentials and asks (“take ₹340 from this card”). Push systems never hand the payer’s credentials to merchants, so credential theft has nothing to eat; fraud migrates to social engineering instead. UPI is push by design; cards are pull. For agents the distinction is existential: a pullable agent is a standing vulnerability, a push-only agent is an auditable actor (cut-0203).

A proposed Ethereum standard for trustless agent identity: on-chain registries through which agents can be discovered, accumulate reputation, and have their work validated across organizational boundaries. Whether or not this specific proposal wins, it names the gap every agentic-payments architecture keeps hitting — payments want a counterparty, and “an agent” is not yet a nameable counterparty anywhere neutral.

The slice of a card transaction paid by the merchant’s bank to the cardholder’s bank, set by network rules. Persistently misread as the price of moving data; better read as a governance fee — it funds the issuer’s fraud losses under liability rules, the dispute apparatus, and the incentive to issue cards at all. Delete it and those costs go unfunded rather than away (cut-0212).

A signed, machine-verifiable record that a principal authorized an agent to make a payment (or class of payments) under stated constraints — amount cap, scope, expiry. The mandate is the evidence layer of agentic payments: when a transaction is disputed, it converts “my agent did what?” from a customer-service scream into a claim that can be ruled on. See The Mandatum for the field guide.

An open standard (originated by Anthropic, 2024) through which AI applications connect to external tools, data sources, and services — a USB port for model context. In payments, MCP matters twice: agents reach payment tools through it, and publications like this one can expose their corpus through it so agents query the source instead of scraping the rendering.

The percentage of each transaction a merchant pays to accept a payment method, split among acquirer, network, and (as interchange) the issuer. India set MDR to zero on UPI and RuPay debit by decree in January 2020 — the natural experiment in what a rail’s protective machinery does when its funding is removed (see row 7 of Plate II).

The entity that legally sells to the buyer — whose name is on the transaction, who owes the taxes, handles refunds, and eats the disputes — regardless of who actually ships the goods. In agentic commerce the MoR question decides everything downstream: when an agent buys through a platform, whether the platform, the merchant, or (nobody wants this) the agent’s operator is the seller of record determines who the regulator calls.

A system shape where many things above and many things below all pass through one thin, stable interface — IP in networking, the UPI switch in Indian payments, possibly a mandate format in agentic payments. The waist’s owner controls the ecosystem’s evolution; the waist’s stability is what lets everything else churn safely. Every serious agentic-payments protocol is, openly or not, competing to become the waist.

Consumption of content where no human ever sees the page: model training, RAG retrieval, agent decision-making. Licensing regimes built on “display to a reader” — ads, paywalls, even much of copyright practice — wobble here. This site’s answer is structural: content is CC-BY with machine twins and embedded citations, so non-display use is invited and attributable, rather than fought and unmeasured.

Charging AI crawlers and agents per request for access to content, rather than serving them free (or blocking them outright) — operationalized by CDN-level schemes and protocols like x402 that quote a price in the HTTP response itself. The first commercial answer to “agents consume the web but don’t see ads.” Its economics are unproven; its architecture — machine-readable prices on machine-consumed resources — is probably permanent.

India’s regulatory category for stored-value instruments — wallets, prepaid cards, vouchers — licensed and supervised by the RBI. The PPI era (roughly 2010–2017) was India’s first mass fintech wave — the world cut-0102’s scar was earned in — and its 2017–18 full-KYC mandate was the sector’s extinction event: the reference case for what happens to lightly-regulated balance-holders at scale.

The moment money actually moves between institutions with finality — as opposed to the confirmation the payer watched, which was a message. Retail rails settle deferred and netted (UPI in batches through RBI accounts; cards days later through network arrangements). Settlement is the money clock’s midnight: everything before it is somebody’s credit exposure (see two clocks).

Any agentic commerce protocol should be adoptable by a corner store in an afternoon — no engineering team, no developer account, no hosted endpoint. UPI passed with a laminated QR code; most agentic protocols currently fail. The test names the difference between winning the head of the merchant distribution and winning commerce (cut-0215).

Every payment system runs an experience clock (what the payer watches: tap, spinner, ✓) and a money clock (what banks obey: clearing, netting, settlement finality, dispute windows). No mainstream rail synchronizes them; the craft is in hiding the gap. Whoever bridges the gap is extending credit and eating risk, named or not. The frame this whole site keeps returning to — see cut-0200 and both Plates.

UPI’s addressing layer: a human-readable handle (name@bank) that resolves inside the network to real account plumbing. The payer holds a stable name; the plumbing can change underneath without the payer noticing. DNS for money (cut-0201) — and the pattern agentic payments has yet to reproduce: a portable way to name who may pay whom, independent of rail.

An open payment protocol (originated by Coinbase, 2025) that revives HTTP status code 402 “Payment Required”: a server quotes a price in a response header, the client — typically an agent — pays, usually in stablecoins, and retries the request with proof of payment. Built for machine-speed micropayments: pay-per-request APIs, pay-per-crawl content, sub-cent prices with no accounts or card forms. Its open question is everything the header omits: disputes, refunds, and who eats a wrong purchase.