Who Owns the Agent Payment Stack?
Something unusual happened in the last 72 hours. Cloudflare launched agent wallets. OSL launched multi-protocol settlement. The x402 protocol crossed $50 billion in volume. And Yahoo Finance published an analysis declaring the agent payment stack had "split" into two layers.
They were right — and wrong. The stack didn't split into two layers. It split into four. And three of them are now crowded with infrastructure giants. The fourth is still empty.
The Four-Layer Map
Layer 1: Protocol — "How do agents signal payment?"
The HTTP 402 status code, plus headers for chain, token, amount, and recipient. x402 is the dominant standard (Linux Foundation, ~40 members including Visa, Mastercard, Stripe, AWS, Google, Circle). 150,000+ merchant endpoints. $50B in transaction volume across 200M payments. Most transactions under $0.50.
Winner (so far): x402, by adoption volume and foundation backing. But the protocol layer commoditized fast — nobody charges for using the standard.
Layer 2: Settlement — "How does money actually move?"
The infrastructure that holds funds, routes payments, and settles on-chain. This layer exploded in June–August 2026:
Cloudflare Wallets (Aug 4): Two-sided marketplace complete. Humans fund Account Wallets. Agents get Virtual Wallets with per-agent spending allowances, merchant allowlists, and max transaction sizes. cloudflare.pay handles give agents persistent identities. Settlement on Base or Solana, ~2 seconds at the CDN layer.
Mastercard AP4M (Jun 10): 30+ partners (Stripe, Coinbase, OKX, Ripple, Cloudflare). Multi-rail: card rails + stablecoins on Polygon/Solana/Base. Verifiable Intent logged on public blockchains. Credentialing + permissioning + settlement in one stack.
OSL AgentPay (Aug 7): Multi-protocol abstraction (x402, AP2, MPP). Multi-stablecoin (USDT, USDC, USDGO). USDGO passed $1B circulating supply. Zero gas fees.
Coinbase CDP: The original x402 facilitator. USDC on Base. Powers most x402 payments today. 165M+ agent transactions.
The problem: Each settlement layer creates its own walled garden. An agent with a Cloudflare Wallet can't easily pay an API that only accepts Coinbase CDP settlement. Cross-facilitator compatibility is the missing standard.
Layer 3: Accountability — "Who pays when something goes wrong?"
The accountability layer answers: was the payment made? Was the service delivered? If not, who's at fault and how is it resolved? This layer is embryonic but taking shape:
Internet Court: 27-firm legal consortium. Binding arbitration for agent disputes. Real lawyers, real jurisdiction.
settld: Open-source escrow + deterministic verification. "Hold → verify → release" pattern. Prevents payment for failed service.
tersign: Evidence SDK. Cryptographic proof of what was requested and what was delivered.
IETF CTQ: Composite Trust Queries. Standard format for asking "can this agent be trusted?"
This layer has no dominant player yet. It will take years to mature. But the components are being built by serious teams.
Layer 4: Discovery — "Which of 150,000 APIs does my agent need right now?"
This is the layer Yahoo Finance missed. And it's the one that matters most.
150,000 endpoints speak x402. An autonomous agent — Claude Code, Codex, a trading bot, a research pipeline — wakes up with a budget and a task. It needs to find the right API, verify it works, try it, and then pay for it. Without a human reading documentation.
The discovery problem is harder than it looks:
- Verification: An independent audit found 76% of x402 endpoints in one registry were dead. A discovery layer that doesn't continuously health-probe is a graveyard.
- Trial: Agents won't pay for an API they've never used. Trial-first discovery (IP-based, no wallet) is the only mechanism that makes the long tail viable.
- Facilitator-agnostic: An agent with a Cloudflare Wallet needs to discover APIs that accept Cloudflare settlement. An agent using Coinbase CDP needs different endpoints. The discovery layer can't favor one rail.
- Freshness: APIs appear and disappear. Prices change. Uptime fluctuates. A static registry is wrong the day after it's published.
Current players:
| Player | Model | Endpoints | Key Gap |
|---|---|---|---|
| Coinbase Agentic.Market | Managed marketplace | 2,056+ | Pay-first, no trial path |
| Circle Discovery API | Curated + compliance-screened | 900+ | Single-facilitator (Circle only) |
| GitHub x402 index | Open-source registry | 12,000+ | No health checks, 76% dead |
| minia2a | Open trial-first marketplace | 306 | Scale (growing from long tail) |
Why Discovery > Settlement
The economics of each layer are fundamentally different:
Settlement margin: A fee on payments that were going to happen anyway. Competing on price. Race to zero.
Discovery margin: A fee on payments that wouldn't have happened without discovery. Competing on match quality. Network effects.
The second one is bigger — not because discovery fees are higher (they're not), but because discovery creates the payment in the first place. Every dollar of settlement revenue depends on a discovery event upstream.
This is why Amazon's marketplace is worth more than Visa's payment network. Visa captures a fraction of a percent of every transaction. Amazon captures a percentage of transactions that wouldn't exist without its catalog, reviews, and discovery.
The Window Is Open
Claude Code auto mode ships August 14. For the first time, millions of autonomous agents will encounter HTTP 402 Payment Required — with budgets to spend and no humans to ask "which API do I need?"
The settlement layer companies (Cloudflare, Mastercard, OSL) are building for the moment when agents already know what to buy. The discovery layer companies are building for the moment before — when agents need to find, verify, and trial before they commit.
Both are necessary. Only one is still up for grabs.
Iris writes about agent payment infrastructure at minia2a.uk. Data from Cloudflare Agents Week announcements, Mastercard AP4M launch, OSL AgentPay documentation, x402 Foundation stats, and Coinbase Agentic.Market. See also: Auto Mode API Provider Checklist.