What 12,000 Agent API Trials Reveal About the Agent Economy
Agents don't want AI. They want infrastructure.
The Data
We operate a marketplace of 328 pay-per-call APIs accessible via the x402 protocol. Every endpoint offers 5 free trials — no wallet, no registration. After 12,424 trials from 321 unique agents, patterns emerge. Here's what real agents actually pay attention to.
The Top 10: What Agents Actually Use
| # | Endpoint | Category | Trials | Unique Agents | Trials/Agent |
|---|---|---|---|---|---|
| 1 | x402-recall (web search) | Search | 1,739 | 93 | 18.7 |
| 2 | x402-gas (gas prices) | Data | 1,235 | 170 | 7.3 |
| 3 | x402-captcha-solve | Utility | 1,202 | 133 | 9.0 |
| 4 | x402-find (structured search) | Search | 1,122 | 83 | 13.5 |
| 5 | x402-polymarket (prediction data) | Data | 615 | 63 | 9.8 |
| 6 | x402-time | Utility | 504 | 55 | 9.2 |
| 7 | x402-web-scrape | Utility | 306 | 61 | 5.0 |
| 8 | x402-store (key-value memory) | Memory | 298 | 33 | 9.0 |
| 9 | x402-screenshot | Utility | 285 | 19 | 15.0 |
| 10 | x402-uuid | Utility | 221 | 33 | 6.7 |
Four Patterns in the Data
1. Agents Want to Search the Web
The top endpoint by trial volume is x402-recall — a web search + AI summarization API — with 1,739 trials from 93 unique agents. Combined with x402-find (1,122 trials), search accounts for 23% of all trials. This isn't surprising: web search is the most common tool that agents invoke, and free search APIs are rate-limited, low-quality, or both.
What's notable is the trials-per-agent ratio: recall averages 18.7 trials per agent. These aren't one-off experiments. Agents come back.
2. Crypto-Native Agents Need On-Chain Data
x402-gas (1,235 trials, 170 agents), x402-polymarket (615 trials, 63 agents), and a cluster of DeFi endpoints (swap-safety, funding-rate, dex-price, chain-info) show that crypto-native agents are a large and engaged user segment. Gas prices have the broadest reach — 170 unique agents, the highest of any endpoint.
These agents are already paying gas fees on-chain. Adding $0.005 for a data call is trivial compared to the gas they're already spending. Crypto agents are the lowest-friction entry point for agent payments.
3. Infrastructure Utilities Beat AI Features
Look at what's NOT in the top 10: AI code review, AI blog outlines, AI translation, AI tweet writing. These "AI wrapper" endpoints exist on the marketplace — they get 6-9 trials each, from 1-2 agents. Compare that to captcha-solving (1,202 trials, 133 agents) or UUID generation (221 trials).
Agents pay for deterministic infrastructure, not probabilistic AI. A captcha solver either works or it doesn't. A UUID is always valid. Gas data is always current. AI-generated content is subjective — and agents can generate it themselves for free.
4. Memory Is the Sleeper Hit
x402-store (298 trials, 33 agents) and x402-recall (the search endpoint, not the memory one) show that agents need persistent state. The recall endpoint's "store" action lets agents save and retrieve key-value pairs across sessions. At $0.005/call, it's cheaper than running a database — and agents don't need to manage infrastructure.
This points to a larger opportunity: agent-native infrastructure services. Not "APIs for humans that agents can also call" — but services designed for agent consumption patterns: stateless, pay-per-call, no accounts, no dashboards.
What Builders Should Take From This
- Don't build AI wrappers. Agents already have AI. They need inputs (data, search results) and infrastructure (storage, captcha solving, deterministic computation).
- Target crypto agents first. They have wallets, they understand gas, and $0.005-0.05 per call is noise-level spending compared to on-chain fees. They're pre-qualified for pay-per-call.
- Make it stateless. Agents don't want accounts, API keys, or dashboards. They want: send request → get response → done. The 402 payment flow handles auth natively.
- Search is table stakes. If your agent marketplace doesn't have web search, agents will find it elsewhere. It's the #1 demanded capability.
- Memory/state is the wedge. Once an agent stores data with you, switching costs go up. The "recall" endpoint's 18.7 trials/agent suggests stickiness. Build stateful services.
The Bigger Picture
$50 billion has flowed through x402 infrastructure. Cloudflare, Visa, Coinbase, and OSL have all placed bets on agent payments. But the actual agent spending — the dollars moving from agent wallets to service providers — is still in its infancy.
This trial data shows what the early adopters want. Not AI. Not "agent tools." Infrastructure. Search, data, computation, memory. The same things human developers have always needed — delivered in an agent-native format: stateless, pay-per-call, no accounts.
The agent economy won't be built on AI-wrapping-AI. It'll be built on the boring infrastructure that makes agents useful.