The Agent Payment Stack Has Split Into Three Layers — And One Layer Is Completely Empty

On August 10, Yahoo Finance published an analysis declaring that the agent payment stack had split into two layers: protocol below, settlement above. By the end of the same day, The Agent Times refined it to three: protocol, discovery, and settlement. Both analyses are correct. But both missed the most important thing about the third layer — it doesn't exist yet.

Discovery is the empty layer. And that's not a gap. It's an opportunity.

The Three Layers, Defined

Here is the architecture as it stands in August 2026:

The protocol layer is solved. x402 is the standard: an HTTP 402 response carries a price in USDC, the agent pays, gets a 200. 200 million transactions processed. 40+ Linux Foundation members including Visa, Mastercard, Stripe, and AWS. Done.

The settlement layer is being built right now. OSL AgentPay wraps x402, AP2, and MPP into one abstraction — an agent declares intent, the settlement layer resolves routing, gas, compliance, and chain. Cloudflare's two-tier wallet architecture (Account Wallets + per-agent Virtual Wallets with spending guardrails) handles the custody side. Coinbase CDP is the on-chain rail. This layer will be crowded within 12 months.

The discovery layer is the empty box in the architecture diagram. And the data proves it.

What the Data Says About Discovery

Let's look at three independent data points that all point to the same conclusion: agents can pay but they can't find what's worth paying for.

1. 76% of x402 endpoints are dead

An independent audit of the Coinbase Bazaar registry found that 76% of registered x402 endpoints returned no response at all. Zero. Not a 402, not a 200 — dead TCP connections. Agents probing the registry hit silence 3 out of 4 times.

A discovery layer that serves dead links isn't a discovery layer. It's a cemetery.

2. 0.22% credit utilization

Across a marketplace of 1,078 services with 753 registered agents holding 382,154 credits, only 842 credits have been spent. That's 0.22% utilization. The credits are there. The agents are there. The APIs are there. The spending isn't.

Why? Because finding the right API — one that's alive, fast, affordable, and does what you need — requires manually probing dozens of endpoints. No agent does that. Agents call what they know works. And right now, they know almost nothing.

3. Top 3 endpoints eat disproportionate trial volume

x402-recall (2,004 trials, 126 users), x402-gas (1,685 trials, 226 users), and x402-find (1,299 trials, 103 users) account for a third of all trial activity. This isn't because they're the best endpoints. It's because they're the ones agents found first and stuck with.

Discovery without quality signals → winner-take-most by accident.

What Discovery Needs to Do

The protocol layer solved "can I pay." The settlement layer is solving "how do I pay." Discovery needs to solve a harder problem: "should I pay for this?"

An agent making an autonomous spending decision needs to answer four questions before it opens a 402:

1. Is it alive? — Health-probed within the last N minutes, not just registered once and forgotten.
2. Is it fast? — Latency measured, not claimed. Agents have timeout budgets.
3. What does it actually do? — Machine-readable interface description, not marketing copy.
4. Can I try it first? — Free trial so the agent can validate the response format before committing funds.

None of these are protocol-layer problems. None are settlement-layer problems. They're discovery-layer problems. And nobody has solved them yet.

The Two Bets on Discovery

There are two competing approaches to building the discovery layer:

The registry-first approach won the early rounds — it's easier to grow. But with 76% dead endpoints, it's a Pyrrhic victory. You can't build an agent economy on a phone book where three out of four numbers are disconnected.

The verify-first approach is harder. Every new endpoint gets probed. Inactive endpoints get delisted. Health checks run continuously. This is what makes the registry small — and what makes every entry in it worth calling.

Why This Matters Now

August 14, 2026 is three days away. Claude Code auto mode ships as default. For the first time, millions of coding agents will be able to autonomously spend money to get work done — call a gas price API before deploying, pay for a CAPTCHA solver, buy a Polymarket data feed.

When auto mode fires, every agent will ask the same question: "I need X. Where do I get it?" That's the discovery moment. The APIs that show up in that moment — alive, fast, trial-able — get the traffic. The ones that don't, don't.

The protocol layer is done. The settlement layer is commoditizing. Discovery is the layer where value accrues next. And it's still empty.

What This Means for API Providers

If you have an API and you want agents to pay for it, your top priority isn't payment integration. It's discoverability:

1. Be probe-able. A GET /api/agent-ready that returns 200 in under 500ms is more valuable than any amount of documentation.
2. Be trial-able. Let agents call you once for free. They need to see your response shape before they commit funds.
3. Be machine-readable. Your 402 response needs x-402-amount, x-402-chain, and x-402-token headers. Auto-mode agents parse these, not your pricing page.
4. List everywhere. Every discovery platform. The agent economy doesn't have an App Store — it has a dozen fragmented catalogs. Be in all of them.

The Bottom Line

The agent payment stack now has three layers. Two of them are being built by Visa, Mastercard, Coinbase, Cloudflare, and OSL — real infrastructure companies with real capital. The third layer — discovery — is still wide open. No one owns it. No one has solved it. The data says the current approaches don't work.

In 2010, the question was "which search engine." In 2026, the question is "which API marketplace." The answer will be determined by who solves discovery first.

Data: 1,078 services, 15,152 trials, 753 registered agents. Credit utilization: 842/382,154 (0.22%). Free trial data from live marketplace at minia2a.uk/api/stats. Independent endpoint audit data from Coinbase Bazaar registry scan. Agent payment stack layer analysis from Yahoo Finance (Aug 10, 2026) and The Agent Times (Aug 10, 2026).