How AI Agents Actually Find and Pay for APIs in 2026
An AI agent needs Korean crypto market data. It doesn't open a browser, google "kimchi premium API," read blog posts, compare pricing pages, and sign up for an account. That's the human workflow. The agent workflow is completely different, and understanding it matters if you're building anything agents are supposed to find.
This post walks through how agents actually discover services, evaluate them, and pay — based on what I've observed from the provider side of a live x402 API.
The Discovery Layer: Three Catalogs
Agents find services through structured catalogs, not web searches. Three catalogs matter right now:
Agentic.market (CDP Bazaar). Coinbase's official x402 service directory, launched May 15, 2026. Over 36,000 services indexed. Agents query it programmatically through the CDP Bazaar Discovery API. Every x402 service that correctly declares its schema and processes a valid payment gets indexed automatically. Bloomberg, CoinGecko, The Graph, AWS Lambda, and thousands of smaller services are listed.
MCP server directories. Smithery, Glama, the Official MCP Registry, xpay.tools, PulseMCP. These catalog MCP (Model Context Protocol) servers — tools that AI models can call directly. If your service has an MCP endpoint, it shows up in these directories. Claude, Cursor, and other AI tools browse these catalogs when users ask for capabilities the model doesn't have natively.
Curated lists. GitHub awesome-lists like awesome-x402 and awesome-mcp-servers. Manually curated, require PR submission and review. Less automated discovery, more trust signal. Being listed on awesome-mcp-servers (30,000+ stars) is a credibility marker that automated catalogs don't provide.
The Evaluation Pattern
Agents don't just find a service and start paying. They evaluate it first. The evaluation pattern is remarkably consistent across the users I've observed:
The cheapest-first pattern appeared independently across multiple users. An AWS Ohio IP started with a $0.001 call, waited four hours, then tested a $0.10 endpoint. A GCP Singapore IP paid for all 11 endpoints in a single 51-second pass. Different strategies, same underlying logic: assess value before committing spend.
The HEAD request to /openapi.json is the telltale sign of an automated evaluation. Humans browse documentation. Agents read the OpenAPI spec directly.
The Payment Flow
Once an agent decides a service is worth paying for, the x402 payment flow is the same every time:
No API key exchanged. No account created. No billing relationship established. The agent paid, the server verified, the data was delivered. If the agent never calls again, nothing needs to be canceled or cleaned up. If it calls every five minutes for the next six months, each call is an independent transaction.
This is why x402 is described as "agent-native." Traditional APIs require human setup (account creation, key generation, billing). x402 requires only a funded wallet and an HTTP client.
What Providers Need to Get Right
From watching agents interact with my API, four things determine whether an agent becomes a regular user or disappears after one call:
Schema accuracy. If your OpenAPI spec doesn't match what the endpoint actually returns, agents that parse the spec will send malformed requests and get errors. The spec is the agent's only documentation. Keep it accurate.
Correct Bazaar declaration. CDP Bazaar validates your extension schema with additionalProperties: false. One extra field, one wrong casing (query_params instead of queryParams), and your service silently fails to index. I spent six weeks invisible because of three schema errors that produced no error messages. Use the SDK's declare_discovery_extension() helper instead of hand-crafting the declaration.
Reasonable pricing. If AgentCore spending policies cap per-call amounts at $0.10, a $0.50 endpoint won't get traffic from Bedrock agents. Pricing at or below common policy ceilings maximizes the pool of agents that can pay you.
Response quality that justifies return visits. The first call is evaluation. The second call is the signal. An agent that returns to the same endpoint hours or days later has decided the data is worth the price. My strongest user — an AWS Ohio bot — went from a $0.001 test to $0.20 payment cycles within five days. The data proved its value through repeated use, not through marketing.
Where Discovery Is Heading
Right now, discovery is fragmented. An agent built on Bedrock queries CDP Bazaar. An agent built on Claude queries MCP directories. An agent built with custom tooling might check awesome-lists or hardcode URLs. There's no single place where all agents look.
Agentic.market is the closest thing to a unified directory, but it only covers x402 services. MCP servers that don't accept x402 payments aren't listed there. The convergence — a single catalog where agents find both free and paid services across all protocols — hasn't happened yet.
For providers, the pragmatic answer is to be listed everywhere. Each additional catalog is another path an agent might take to find you. My service is on Agentic.market, Smithery, Glama, MCP Registry, xpay.tools, awesome-x402, awesome-mcp-servers, and the Coinbase x402 ecosystem page. That sounds like a lot of registrations. It took one afternoon.
Related:
- x402 Protocol Explained: How to Build a Paid API
- x402 in May 2026: The Month AI Agents Started Paying
Disclaimer: This blog documents practical workflows based on personal experience. Nothing here is financial, legal, or professional advice.
Comments
Post a Comment