Glossary
What is agentic commerce?
Agentic commerce is what happens when AI agents — not humans — discover services, negotiate price, and complete transactions on their own. The buyer is software. The seller is often software too. Settlement happens in milliseconds, without a checkout page.
The shift in who's doing the spending
For the entire history of e-commerce, the payment model has assumed a human at the keyboard. Click “buy.” Type a card number. Approve a recurring charge. Every checkout flow, fraud rule, and dispute system was built around that assumption.
Agents don't fit. An agent solving a research task might query 40 paid APIs in 30 seconds, each charging fractions of a cent. An agent booking travel might negotiate with three different vendor agents before settling. An agent running a metered SaaS workflow might spawn a sub-agent that needs its own scoped budget. The UX surface that worked for humans — popups, redirects, re-authorization prompts — actively breaks at this scale.
Three problems agentic commerce has to solve
- Identity. When agent A buys from agent B, both sides need to know who they're dealing with — a stable, verifiable identity that survives across sessions and platforms. Stripe's customer record doesn't generalize; emerging standards like ERC-8004 do.
- Metering. The value exchanged is per-call, per-token, or per-result. You need to attribute every fraction of a cent to the right buyer and seller in real time, and you need that attribution to be tamper-proof for downstream auditing. More on metered payment models.
- Payments. The actual money has to move. Sometimes that's a Visa virtual card on real card rails; sometimes it's USDC; sometimes it's off-chain credits with periodic on-chain settlement. The agent shouldn't have to care which.
The protocol layer
Several open standards are converging to make agentic commerce work. None of them is the “winner” yet, and the bet most builders make is that more than one will persist.
- x402 — HTTP-native payment handshake. The wire protocol that turns “here is a paywall” into “here is a paid response” in one round trip.
- A2A — Agent-to-Agent protocol. The conversation layer between two agents that lets them discover each other's capabilities before exchanging value.
- MCP — Model Context Protocol. Lets an LLM expose tools that other agents can call, with x402 as the payment carrier.
- AP2 — an emerging agent-payment specification (cited alongside MCP / A2A / x402 in industry roadmaps). Specifics are still settling; track the spec for details.
Nevermined's position is rail-agnostic: whichever of these protocols you adopt, the metering, identity, and settlement primitives are the same. See how the SDK abstracts the protocol layer.
Examples in the wild today
- A research agent that pays per-query to access primary sources and gets reimbursed by the user at the end of the session, plus a markup for the orchestration.
- A travel-booking agent that negotiates with hotel and airline agents in parallel, settles on a multi-leg itinerary, and charges the user's delegated card.
- A coding agent that consumes paid linting, security scanning, and documentation-generation tools per pull request, with the cost amortized into the user's subscription.
- An MCP-based OpenClaw plugin that runs in a chat channel, taking payment for plans and per-message tools without a separate checkout experience.