What is agentic commerce.
Agentic commerce is commerce conducted by AI agents acting on behalf of humans or businesses. An AI agent might buy an API call, pay a microservice, hire another agent, or receive payment for work it produces. Agentic commerce requires three things: a payment identity per agent, programmatic spending controls, and APIs that return machine-readable payment requirements rather than human checkout flows.
The traditional checkout assumes a human is at the keyboard.
For thirty years, the dominant payment pattern on the web has been the checkout flow: a buyer arrives, signs up, enters card details, clicks Pay. Every layer of that pattern (the form, the 3DS challenge, the verification email, the saved card) assumes a human is making the decision in real time. AI agents cannot complete any of those steps. They cannot fill in forms, cannot reliably remember passwords, cannot click confirmation emails on a delay, cannot pass card-network risk scoring optimized for human behavior.
As more software calls are made by AI agents instead of humans (an irreversible trend now that agent frameworks like LangChain, CrewAI, OpenAI Agents SDK, and AutoGen have matured), commerce needs payment rails that work without a human in the loop. Stablecoin rails on cheap chains fit this constraint; card rails do not. Programmatic 402 Payment Required responses fit; HTML signup forms do not. Agentic commerce is the name for that shift.
The economic implications are larger than they look. If even 5% of API calls in five years are made by agents rather than humans, and each agent-driven call needs a payment rail that works programmatically, the share of internet commerce flowing through agentic-commerce infrastructure could rival the share flowing through Stripe today.
Four structural differences.
- The buyer is an agent, not a human. The buyer cannot read marketing copy, evaluate trust signals visually, or fill out forms. Conversion-rate optimization that works on humans (above-the-fold CTAs, trust badges, urgency timers) does nothing for agents. What matters is whether the agent's authorized spend policy permits the purchase and whether the API returns machine-readable payment requirements.
- Payments are smaller and more frequent. Per-call billing replaces subscriptions for many use cases. A coding agent might make 200 paid calls in an afternoon at $0.02-$0.10 each. The economics of card rails fall apart at this transaction shape (interchange fees and minimum transaction sizes); stablecoin rails on cheap chains work.
- Stablecoins fit better than cards. Chargebacks are not meaningful between agents (the threat model is different from human-to-merchant disputes). Settlement finality matters more than reversibility. Cross-border payments cost the same as same-country ones. Cards were not designed for this shape; stablecoins effectively were.
- The seller returns 402 Payment Required. Instead of a checkout form, the seller's API returns a 402 describing how to pay (network, asset, amount, recipient). An x402-aware client reads it, pays on-chain, and retries the call programmatically - no human in the loop. The hosted-checkout variant (a payment link a person can click) exists too, for buyers who are not yet x402-aware.
What agentic commerce looks like in practice.
Four concrete patterns from the field, ordered roughly by how common they are today.
An LLM coding assistant paying per call
Your IDE-integrated coding agent calls a code-search API every time it needs context. Each call costs $0.05 in USDC on Base. The agent never asks you to approve a payment; the pre-authorized spend policy on its wallet covers calls up to $20/day. Over a month, the agent buys $300 of code-search lookups.
A research agent paying a competitor's premium API
A research agent assembling a market report needs one paid query from a competing dataset to validate a claim. It calls the competitor's API, gets a 402 Payment Required, parses the hosted URL, pays $5 USDC, retries the call. The competitor's bookkeeping sees a programmatic sale; the agent's bookkeeping sees a logged $5 outgoing payment.
An MCP server charging per tool invocation
A semantic-search MCP server exposes its tool to Claude Desktop, Cursor, and Cline. Each tool call costs $0.02. Over an hour-long debugging session, the user's coding agent invokes the MCP tool 35 times and pays $0.70 in aggregate. No subscription, no signup form - the payments happen as the work happens.
An agent paying another agent for routing
A primary research agent delegates a section of work to a specialist agent (translation, fact-checking, formatting). The specialist agent has its own pricing page; the primary pays the specialist's listed rate in USDC. From the user's perspective, they paid one fee to the primary agent; the primary handles paying its delegate.
Pair this entry with these.
Agentic commerce is the umbrella concept. These three terms cover the layers that compose it.