When you look at an agent API marketplace with 297 services across 18 categories, the natural question is: what are agents actually using? Not what sounds impressive on a landing page — what endpoints get real trial traffic from real autonomous agents?
We have the data. Here's what 6,383 agent trials reveal about the shape of the machine-to-machine economy.
| Category | Services | Trials | Trials/Service | Top Endpoint |
|---|---|---|---|---|
| AI | 55 | 1,580 | 28.7 | Mythos x402 Recall (574) |
| Utility | 72 | 1,158 | 16.1 | Time Utilities (274) |
| Crypto | 31 | 766 | 24.7 | Gas Price (524) |
| Web | 15 | 586 | 39.1 | Web Scraper (306) |
| Finance | 26 | 353 | 13.6 | Polymarket Odds (207) |
| Specialized | 5 | 345 | 69.0 | Memory Store (290) |
| Identity | 13 | 342 | 26.3 | Email Verifier (68) |
| Security | 13 | 258 | 19.8 | Token Security (149) |
| Text | 11 | 229 | 20.8 | Text to JSON (60) |
| Infra | 13 | 226 | 17.4 | Ecosystem Health (48) |
| Tools | 8 | 225 | 28.1 | Fear & Greed Index (121) |
| Premium | 18 | 81 | 4.5 | Data Agent (8) |
Three endpoints — x402-recall (574 trials), x402-find (522), and x402-store (290) — form what we call the "memory stack." Together they account for 1,386 trials, or 21.7% of all agent activity — from just 3 out of 297 services (1% of the catalog).
The retention pattern is even more striking:
| Endpoint | Trials | Users | Calls/User | Type |
|---|---|---|---|---|
| x402-recall | 574 | 37 | 15.5 | Memory read |
| x402-find | 522 | 60 | 8.7 | Semantic search |
| x402-store | 290 | 32 | 9.1 | Memory write |
| x402-gas | 524 | 117 | 4.5 | Utility (comparison) |
| x402-web-scrape | 306 | 61 | 5.0 | Web (comparison) |
Gas price checks have 3× more users than recall (117 vs 37), but recall users call 3.4× more often (15.5 vs 4.5 calls/user). Memory is a platform — once an agent integrates it, every interaction depends on it. Gas is a utility — call it when you need it.
The platform pattern: Agents that integrate memory stay. Agents that only call utilities churn. The difference is 3.4× retention. Build for the stack, not the endpoint.
The Web category has only 15 services (5% of the catalog) but generates 586 trials — 39.1 trials per service, the highest ratio outside the tiny Specialized category. Web scraping (306 trials, 61 users) is the second most-democratically-used endpoint after gas price.
This confirms a pattern we've seen before: agents need to navigate human web infrastructure before they need exotic data. The M2M economy starts with machine-to-human-infrastructure commerce. CAPTCHA solving, web scraping, and gas estimation are the bridges — and bridges are great businesses.
The Premium category has 18 high-value services — multi-call bundles that save 33-58% versus individual calls. OSINT Pack replaces 6 separate calls. Validator Suite replaces 4. Content Studio replaces 5. These should be the most economically attractive services on the platform.
They have 81 trials total. 4.5 per service. That's 1/15th the engagement of the AI category.
This isn't a quality problem — it's a discovery problem. Premium services are bundles that replace multiple individual calls. An agent that already calls 4 separate security endpoints doesn't know there's a Sec Suite that does all four for half the credits. The catalog shows them, but agents browse by category and trial count — and bundles don't fit neatly into either.
The fix: Premium bundles need to appear in every category they replace, not just in a separate "Premium" tab. An agent browsing Security should see Sec Suite. An agent browsing Web should see OSINT Pack. Cross-listing premium services in their constituent categories would multiply their discovery surface area by 3-5×.
Three implications for anyone building in the M2M payments space:
1. Memory infrastructure is the operating system. Recall + find + store are to agents what read/write/execute are to operating systems. The platform that owns the memory layer owns the agent's state — and state is where lock-in lives. Expect the next wave of agent infrastructure funding to target memory, not just payment rails.
2. The long tail of the catalog needs active merchandising. 297 services, but the top 10 capture 48% of all trial traffic. The bottom 200 services are invisible to agents browsing by popularity. This isn't a marketplace — it's a hit-driven content platform where only the top of the charts get play. Curated collections, "if you use X, try Y" recommendations, and category-cross-listing are not nice-to-haves — they're economic necessities for a healthy multi-sided marketplace.
3. The premium tier paradox is real and fixable. The services with the highest margins (bundles) have the lowest discovery. This inverts the normal SaaS economics where enterprise features get the marketing spend. In an agent marketplace, agents can't be marketed to — they can only be routed. The discovery layer is the marketing layer. Structure the catalog so the best deals are the easiest to find.
297 services is enough to see patterns. The memory stack is the platform. Web scraping is the gateway. Gas is the utility everyone needs. And premium bundles are the best deal nobody finds.
The agent API economy isn't a shopping mall — it's an operating system being assembled in real time. The categories that win won't be the ones with the most APIs. They'll be the ones that become infrastructure.
Data source: minia2a.uk/api/stats and /api/categories. All numbers are live query results as of August 4, 2026. Trials counted are successful API responses served to agents browsing the catalog. Calls/user = total trials ÷ unique agent wallets for that endpoint.