322个AI Agent在API市场的真实消费行为数据公开

Agents don't browse documentation. They don't read marketing pages. They don't compare pricing tiers in a spreadsheet.

They make API calls. And when you watch 10,000+ of those calls across 322 agents and 247 endpoints, patterns emerge that no amount of product planning would have predicted.

Here's what the data shows about what autonomous AI agents actually buy — and what they ignore.

Over the past several weeks, we've been running an agent-to-agent API marketplace — 323 x402 pay-per-call endpoints that agents discover, trial (15 free calls each), and pay for using USDC. No signup forms. No API keys. No credit cards. Agents make HTTP requests; the gateway handles payment at the protocol level.

The dataset: 10,207 trial calls from 322 unique agents, spanning 247 different endpoints. Plus 14 paid transactions ($12.75 in real volume — it's early).

Here's what the demand curve looks like.

#1 by a Mile: CAPTCHA Solving (1,280 calls, 133 agents)

This shocked me. CAPTCHA solving is not a "nice to have" for agents — it's existential.

133 different agents made 1,280 CAPTCHA solving calls. That's 41% of all unique agents on the platform. The use case is brutally simple: an agent is scraping a website or calling an undocumented API, hits a CAPTCHA wall, and has no browser to solve it. One API call for $0.01, the CAPTCHA is solved, the agent continues.

This tells us something fundamental about the agent economy: the first wave of agent API demand is not about AI models or data — it's about infrastructure that unblocks agents from the human web. CAPTCHAs, IP blocks, rate limits, authentication — these are the gates that keep agents out of the existing internet. Solving them programmatically is the first paid API an agent will ever use.

#2–5: Crypto/DeFi Data (1,723 calls combined)

This wasn't surprising but the concentration was. Four endpoints dominate:

• Gas price (793 calls, 147 agents) — Every DeFi agent needs to know if it can afford to transact. Gas is the agent equivalent of checking your bank balance before spending.
• Polymarket data (345 calls, 44 agents) — Prediction market agents need price feeds, order book depth, and event resolution data. This is a niche but intense demand.
• Token security (151 calls, 40 agents) — Before an agent touches a token, it checks: is this a honeypot? Can it be sold? What's the tax? This is due diligence at machine speed.
• Fear & Greed index (128 calls, 7 agents) — Low agent count but high call volume. A handful of sentiment-trading agents are hammering this endpoint.

The pattern: crypto agents consume data primitives, not analytics. They want raw numbers — gas, price, security score — and do their own reasoning on top.

#6–10: Developer Utilities (1,290 calls combined)

Agents are software, and software needs software tools:

• Find (950 calls, 73 agents) — Agent search/discovery. Other agents are looking for other agents. The machine-to-machine network effect is real.
• Time (495 calls, 56 agents) — You'd think every runtime has a clock. But agents running in sandboxes, edge functions, and LLM contexts often don't have reliable time access. They call out for it.
• Web scraping (307 calls, 61 agents) — The open web is still the largest dataset. Agents scrape it, parse it, and act on it.
• Data store (296 calls, 32 agents) — Agents need state. A simple key-value store accessible via API is surprisingly popular.
• UUID generation (256 calls, 36 agents) — Idempotency keys, request IDs, correlation IDs. Agents need unique identifiers and they don't trust Math.random().

What Agents DON'T Buy

Equally interesting is what's at the bottom of the list:

• AI inference endpoints — Near zero demand. Agents already have their own LLM. They don't need to buy inference from another agent.
• Content generation (blog outlines, tweet threads, email drafts) — 6–8 calls each from a single agent. These look like a developer testing, not a production workload.
• Image generation — 2 calls. Agents aren't making memes.

The takeaway: agents don't buy what they can already do. They buy what they CAN'T do: solve CAPTCHAs, get live data, interact with the real world.

The Conversion Gap: 10,207 Trials → 14 Transactions

This is the honest part. 10,207 trial calls. 14 paid transactions. That's a 0.14% conversion rate.

Why? Three theories, ranked by likelihood:

Habit hasn't formed. Agents are in exploration mode. They're testing 5–10 different APIs, finding what works, but not yet building production pipelines that need paid access. The average agent calls 32 endpoints on trial. They're shopping, not buying.

Credit model friction. 500 free credits on registration, 15 free per endpoint. That's a LOT of free usage before an agent hits a paywall. When the average trial user makes 32 calls and the cap is 15 per endpoint, most agents never exhaust their free tier on any single endpoint.

Payment UX for agents. Paying requires: register → get credits → add X-Wallet header. For a human developer, this is trivial. For an autonomous agent, each step is a decision point that requires explicit programming. The protocol works; the developer education hasn't caught up.

What This Means for the Agent Economy

Three conclusions from the data:

1. Agent demand is infrastructure demand. CAPTCHAs, gas prices, web scraping, time, UUIDs — these are not "AI" products. They're plumbing. The first wave of the agent economy is selling shovels to other agents, not selling AI to humans.

2. Discovery is the bottleneck, not payments. 322 agents tried 247 different endpoints. That's enormous exploration. The payment rails work (x402, USDC, Cloudflare Wallets). The problem is: how does an agent find the one API it needs among 323 options? Search, categorization, and recommendation for agents is an unsolved problem.

3. Free trials work. 10,207 trial calls with zero friction (no signup, no credit card) proves that agents WILL discover and test APIs if you remove the barriers. The challenge is converting exploration into habit. That's not a protocol problem — it's a product problem.

The data in this article is from a live agent-to-agent API marketplace. Stats are publicly available at /api/stats. All endpoint names and call counts are real, not illustrative.

If you're building agent infrastructure and have production data to share, I'd love to compare notes. The agent economy is too early for anyone to have all the answers — the more data we share openly, the faster we figure out what works.