The agent payment rails are built. x402 has processed $50B across 200M transactions. Cloudflare launched per-agent wallets on August 4. OSL AgentPay launched multi-stablecoin settlement on August 7. Coinbase opened its agent payment stack to all business clients. The Linux Foundation's x402 Foundation counts ~40 members including Visa, Mastercard, Stripe, AWS, and Google.
The question is no longer "can agents pay?" — it's "which of 150,000 merchant endpoints does my agent actually need?"
This is the discovery problem. And August 2026 is the month it became the bottleneck.
The Four Models of Agent API Discovery
Four competing approaches are emerging to answer the same question: how does an AI agent find the right API, verify it works, and pay for it?
| Model | Example | How It Works | Strength | Risk |
|---|---|---|---|---|
| Curated | Circle Discovery API | 900+ services, compliance-screened, continuously health-checked. Dead endpoints auto-excluded. | Quality guarantee — every endpoint works | Single-issuer lock-in (USDC on Circle rails) |
| Marketplace | Coinbase Agentic.Market | 165M+ txns, 480K+ agents, thousands of services. Pay-first model with x402 + USDC on Base. | Scale — the 800lb gorilla with network effects | Pay-first: no trial means long-tail endpoints are invisible |
| Trial-First | minia2a.uk | 328 services, 5 free calls per endpoint (anonymous IP), 500 free credits on registration. No wallet required to try. | Discovery without commitment — niche endpoints get discovered | Scale (328 vs 900+) — but growing |
| Registry | x402 Discovery Index (GitHub) | 12,000+ endpoints in an open-source JSON index. Anyone can list, anyone can query. | Truly open — no gatekeeper | No verification — 76% of endpoints in one recent audit were dead |
Why Trial-First Matters More Than Ever
The curated model (Circle) and the pay-first model (Coinbase) both solve for quality. They're great for the top 200 endpoints — the ones agents already know they need. Gas prices, token security, web scraping.
But an independent audit published last week found that 76% of endpoints in open x402 registries are unresponsive. These endpoints — built by solo developers, niche data providers, experimenters — disappear silently. In a pay-first or curated model, they simply vanish from the agent's view.
Trial-first is the only model that makes the long tail viable:
- Builders can list an endpoint and get real usage data (trial counts, user IPs, error rates) before anyone commits money.
- Agents can test an endpoint's accuracy, latency, and schema before adding it to a workflow.
- The marketplace gets organic quality signals — endpoints that get retried (not just tried once) float to the top.
Real-time stats from /api/stats as of August 9, 2026. 252 of 328 endpoints have received at least one trial call — meaning 77% of listed services have been discovered by at least one agent.
The Gap Nobody's Talking About: Credit Activation
Here's a number that keeps me up: across minia2a, 34,597 free credits have been issued but only 787 have been spent (2.3%). 60 developers created wallets, claimed their 500 credits, and… stopped.
This isn't a minia2a-specific problem. It's an industry pattern:
- Agent developer registers — excited about the concept of agents paying agents.
- Claims free credits — 500 credits is enough for 100-500 API calls depending on the endpoint.
- Goes back to building — the credits sit idle while they work on their agent's core logic.
- Never returns — by the time their agent can make API calls, they've forgotten about the credits or moved to a different approach.
The fix isn't more credits. It's embedding the spend moment into the development workflow. When an agent framework (Claude Code, Codex, CrewAI, AutoGen) encounters a 402 response, the developer should see: "This API costs 2 credits. You have 500. Spend now? y/n" — not a payment wall.
What minia2a Can Do That Others Can't
Each discovery model has a blind spot. Here's where trial-first wins:
- Verify-first listing — Every endpoint on minia2a is probed on registration. Dead endpoints are flagged, not silently removed (transparency) and not silently kept (curation). Builders get notified: "Your endpoint returned 500 — fix it or it goes inactive."
- Facilitator-agnostic routing — minia2a doesn't care whether the payment settles via Coinbase CDP, Cloudflare Wallets, or a self-hosted facilitator. The marketplace routes to whatever the endpoint supports. Circle and Coinbase marketplaces route through their own rails only.
- Anonymous trial → wallet graduation — The first 5 calls per endpoint require nothing. No wallet, no registration, no gas. This is the only model where an agent can discover, test, and decide before committing any identity or funds.
The Bottom Line
The agent payment infrastructure is converging: x402 is the standard, USDC is the currency, and Cloudflare/Coinbase/OSL are the rails. The differentiation is moving up the stack to discovery and trust.
Curated marketplaces will capture the top 200 endpoints. Open registries will capture the long tail. Trial-first marketplaces sit at the intersection — making the long tail discoverable by letting agents try before they pay.
The model that wins isn't the one with the most endpoints. It's the one where an agent can say: "I need X. Show me endpoints that do X. Let me try the top 3. Pick the best one. Pay."
The payment rails are built. Now we build the brain.