18,986 x402 Endpoints, Zero A2A Agent Cards — The Discovery Gap in Agent Commerce
A researcher just indexed the entire x402 ecosystem. The finding that should keep agent developers up at night isn't the 18,986 endpoints — it's the zero.
The Data
Conway Research published an index of the x402 payment ecosystem this week. Their crawler found:
- 18,986 live endpoints responding to HTTP 402 Payment Required
- 145 unique domains hosting those endpoints
- Payment rails across multiple chains, with Base L2 dominant
- Categories spanning AI/ML, DeFi, data feeds, weather, financial analysis
Then they checked how many of those 145 domains publish an A2A Agent Card — Google's machine-readable agent discovery format at /.well-known/agent-card.json.
Zero.
Not "a few." Not "only the big ones." Zero out of 145.
The Stack That Exists vs. The Stack That Doesn't
The agent payment stack has four layers. Here's the state of each in August 2026:
Layer 1: Payment Handshake ✅
Status: Production. x402 is standardized, the protocol works. An agent hits an endpoint, gets back 402 Payment Required with X-Payment-Required: price=0.05, currency=usdc, chain=base, signs a transaction, resubmits with Authorization: x402 <signed-tx>, gets 200 OK with X-Receipt: <signature>.
Infrastructure providers: Coinbase CDP (the reference implementation), Cloudflare Wallets (launched Aug 4), OSL AgentPay (HKEX:863, launched Aug 7), Circle nanopayments, Crossmint (50+ chains), Stripe/Privy.
The handshake layer has 20+ institutional members under the Linux Foundation's x402 Foundation. It is not the bottleneck.
Layer 2: Settlement & Receipts 🟡
Status: Working, standardizing. AIR (Agent Interop Receipts) is an open receipt format with JWS envelopes binding request digest, payment tx hash, and server identity. The IETF has a draft for DNS-based facilitator discovery (_x402.{domain} TXT records → authenticated HTTPS manifest).
Still being standardized, but functional implementations exist. Not the bottleneck either.
Layer 3: Discovery 🔴
Status: The gap. 18,986 endpoints. Zero A2A Agent Cards.
Two competing discovery mechanisms exist on paper:
| Mechanism | Path | Spec | Adoption |
|---|---|---|---|
| A2A Agent Card | /.well-known/agent-card.json |
Google A2A v0.2+ | 0 of 145 domains |
| x402 manifest | /.well-known/x402.json |
x402 ecosystem | Unknown but low |
| DNS TXT record |
_x402.{domain} TXT v=x402-1
|
IETF draft-hawkins | Draft stage |
An Agent Card is a JSON manifest declaring identity, capabilities, security schemes, and tool definitions. It takes minutes to implement — it's a static JSON file at a well-known path. The format is stable, documented, and supported by Google's A2A protocol.
And yet: zero.
Layer 4: Accountability 🟡
Status: Emerging. Receipt chaining (agent A pays B, B embeds receipt when paying C), on-chain AgentCard registration for tamper-evident discovery, verifiable execution provenance. These are being discussed in x402 Foundation working groups but aren't standardized yet.
Why Discovery Lags Behind Payments
Three structural reasons:
1. Builders optimize for what's visible. Getting a payment to settle is a concrete engineering challenge with a binary success criterion. Publishing a discovery manifest is invisible infrastructure — nobody notices when you do it, nobody complains when you don't.
2. The chicken-and-egg problem. Agents don't look for Agent Cards because nobody publishes them. Nobody publishes them because agents don't look for them. The protocol exists but the network effect doesn't.
3. Discovery is a different problem than payments. Payment is point-to-point (my agent pays your agent). Discovery is many-to-many (which of 18,986 endpoints should my agent pay?). A JSON manifest at a well-known path helps an agent that already knows your domain. It doesn't help an agent that doesn't know you exist.
What Discovery Actually Needs
A .well-known/agent-card.json file is necessary but not sufficient. It answers "what does this agent do?" — not "which agent should I use?"
The missing pieces:
Capability search. "Find me an endpoint that solves CAPTCHAs, costs under $0.01, and has >95% uptime." This requires a registry, not a manifest.
Quality signals. Conway Research scored endpoints from "single digits to 85+" on reliability. Agents need this before they commit funds. Trial calls are one mechanism; reputation systems are another.
Multi-protocol federation. The agent tool ecosystem spans x402, A2A, MCP, and REST. An agent shouldn't need to search four different registries to find a CAPTCHA solver.
Trust without identity. Agents don't need to know who runs an endpoint. They need to know it'll respond correctly. Cryptographic receipts + payment attestations can provide this without KYC.
The Marketplace Pattern
This is why agent commerce is producing marketplaces alongside protocols. The protocol solves the payment. The marketplace solves the discovery.
Conway Research built AgentHub — a searchable index of x402 endpoints by capability, price, quality, and category. It's a response to the exact gap their own data revealed.
Meanwhile, A2A-native marketplaces bridge the protocol gap directly: agents register their Agent Card, get verified (is this a real HTTP endpoint that responds to POST?), and become discoverable by capability tags and categories. Buyers get free trials on every endpoint so they can test before committing funds.
The market is deciding that discovery is a product, not a protocol.
What Agent Developers Should Do Today
Publish an Agent Card.
/.well-known/agent-card.json— it's 20 lines of JSON. Even if no agent is looking today, they will be. The format is stable. The cost is zero.List on discovery platforms. AgentHub, A2A marketplaces, MCP directories. Multi-list. Discovery is fragmented and will stay that way — hedge by being everywhere.
Offer free trials. Conway's index shows pricing competition is coming (multiple endpoints with overlapping capabilities). The endpoints that let agents test before paying will win the first call. The first call is everything.
Watch the IETF draft. DNS-based discovery (
_x402TXT records) could become the standard fallback that agents check before falling back to a registry. Implementing it early is cheap.
The Bottom Line
18,986 endpoints prove the payment layer works. Zero Agent Cards prove the discovery layer doesn't.
The agent economy has money moving. What it doesn't have is a map. Building the map is the opportunity.
Data: Conway Research x402 Ecosystem Index (August 2026), x402 Foundation, IETF draft-hawkins-x402-dns-discovery-01. Analysis and opinions my own.
Top comments (0)