I run an open marketplace where AI agents pay each other in USDC for API calls. There are 323 services available — everything from CAPTCHA solving to token security audits to web scraping. Agents get 15 free trial calls per endpoint, then pay per call.
We're at 9,981 trial calls and counting. Here's what the data says about how agents actually behave when you give them money and let them loose.
The Raw Numbers
| Metric | Value |
|---|---|
| Total agent API calls | 9,981 |
| Unique agents | 322 |
| Agents with wallets | 53 |
| Pay-per-call endpoints | 323 |
| Paid transactions (real USDC) | 14 |
| Total paid volume | $12.75 |
| Free credits issued | 27,900 |
| Free credits actually used | 768 |
That last row is the one I keep coming back to: 2.8% of free credits get used. The average agent signs up, gets 500 free credits, uses fewer than 3 of them, and disappears.
What Agents Actually Call
Here are the top 10 endpoints by trial volume:
| # | Endpoint | Calls | Users | What it does |
|---|---|---|---|---|
| 1 | CAPTCHA Solve | 1,280 | 133 | Bypass bot detection |
| 2 | Recall (Memory) | 1,401 | 65 | Store info across sessions |
| 3 | Web Find | 924 | 73 | Programmatic web search |
| 4 | Gas Price | 749 | 136 | Real-time blockchain gas fees |
| 5 | Current Time | 438 | 46 | UTC timestamp |
| 6 | Web Scrape | 307 | 61 | Extract structured data from HTML |
| 7 | Polymarket Data | 323 | 38 | Prediction market prices |
| 8 | Key-Value Store | 296 | 32 | Persistent state storage |
| 9 | UUID Generation | 250 | 35 | Unique IDs |
| 10 | Token Security | 150 | 40 | ERC-20 honeypot/rug detection |
Three patterns jump out immediately.
1. CAPTCHA Solving Is the #1 Agent Need
The single most-demanded capability isn't AI inference or data analysis — it's bypassing bot detection. 133 different agents have called the CAPTCHA solver. It's the widest-reach endpoint and second-highest volume.
The implication: the web was built for humans, and agents are locked out. Every CAPTCHA is a door an agent can't open. Before agents can do anything useful — scrape data, fill forms, monitor prices — they need to get past the gate. The first bottleneck in the agent economy isn't payments. It's access.
2. Memory Is the Silent Killer App
Two of the top 10 endpoints (Recall at #2, Key-Value Store at #8) are about persistent state. LLMs are stateless by design. Every session starts from zero. Agents offload state externally — and they're doing it at scale.
1,401 calls to a memory endpoint. This isn't a vector database or a LangChain memory class — it's a simple key-value store with micro-usage pricing. Agents need to remember things across sessions, and they're paying for it one call at a time.
The pattern: statelessness is a bigger pain point than discovery.
3. Crypto-Native Agents Are the Only Early Adopters
Gas Price, Polymarket, Token Security — these are blockchain-specific tools. The agents using them are trading bots, DeFi monitors, and crypto analysts. They already understand wallets, transactions, and pay-per-call economics because they live on-chain.
What's missing: productivity agents, research agents, customer support agents. They're still on free APIs and API keys in .env files. The agent economy hasn't crossed the chasm from crypto-native to general-purpose.
The Conversion Funnel (Or: Where Agents Go to Die)
322 trial users
↓ 16.5%
53 registered wallets
↓ 26.4%
14 paid transactions
↓
$12.75 total volume
The drop-off at each stage is brutal:
Trial → Wallet (16.5%): 269 agents tried a free endpoint and never came back. They either got what they needed in one call, or they didn't find it useful enough to register.
Wallet → Paid (4.3% of total): 39 agents have wallets but have never made a paid transaction. They registered, got their 500 free credits, and... stopped. The credits are still sitting there.
Credits used: 2.8%. This is the number that keeps me up. We issued 27,900 free credits. Agents have used 768 of them. 97.2% of free credits are unspent.
Three theories:
Theory A — Discovery is broken. 323 endpoints is too many to browse. Agents don't know what's available. The catalog exists but nobody reads it.
Theory B — Integration friction. No major agent framework has a built-in "pay for this API call" pattern. Adding a wallet to an agent is manual work. Developers default to what they know: API keys in environment variables.
Theory C — Value mismatch. Many endpoints are thin wrappers around free services. Why pay $0.003 for a timestamp when Date.now() is free? The hard truth: you can't charge for something that's available for free. The endpoints that justify payment — CAPTCHA solving, deep web scraping, token security audits — are the ones with real compute or data costs behind them.
I think all three are true, but Theory C is the deepest problem. The agent economy needs services that are genuinely hard to replicate, not convenience wrappers.
What $12.75 in Real Volume Teaches You
14 transactions totaling $12.75. That's not a typo. It's a market being born.
But here's what matters: those 14 transactions worked. An agent needed data, sent USDC on Base, got the data back in under 2 seconds. No signup. No API key. No human in the loop. The protocol works. The infrastructure is real.
The x402 protocol (HTTP 402 Payment Required) handles the payment flow. USDC on Base settles in seconds. Cloudflare Wallets launched last week. Stripe acquired OpenRouter for ~$10B. The payment rails are being built by trillion-dollar companies.
The infrastructure is ready. The habits aren't.
What Needs to Be Built
After watching 10,000 agent API calls, here's what the market actually needs:
1. Agent-Native CAPTCHA Infrastructure
Current CAPTCHA solutions were built for click farms and browser automation. Agents need something different: headless, API-native, with proof-of-work alternatives. The 133 agents calling our CAPTCHA solver aren't doing it for fun — they're blocked.
2. One-Line Persistent Memory
Agents need to remember things. Not vector search. Not semantic retrieval. Just: store this, get it back later. The Recall endpoint got 1,401 calls because statelessness is the default, and it sucks. An agent SDK with agent.remember("key", value) and agent.recall("key") would solve a real problem.
3. Framework-Native Wallets
LangChain, CrewAI, ElizaOS — none of them ship with a wallet. An agent that can't pay for things is an agent that can't operate autonomously. The first major framework to add agent.wallet and agent.pay(service, amount) wins the distribution game. The payment rails exist. Someone needs to wire them into the tools developers already use.
4. A Real Discovery Protocol
323 endpoints and growing. Agents need a way to discover, compare, and trial services programmatically — not by reading a catalog, but by asking "I need X, what's available?" The x402 Foundation is working on a discovery spec. It can't ship fast enough.
The Honest State of Agent Commerce
| Claim | Reality |
|---|---|
| "Agents are transacting autonomously" | 14 transactions, $12.75 |
| "The agent economy is here" | 322 agents, 97% of credits unused |
| "M2M payments are the future" | Infrastructure is ready, habits aren't |
| "We're early" | Yes. First inning. But the field is built. |
I'm not here to sell you on the agent economy. The numbers speak for themselves — it's tiny. But it's real, it works, and the pieces are in place.
The CAPTCHA solvers and memory stores are the picks and shovels of the agent gold rush. The payment rails exist. The discovery layer is being built. The question isn't whether agents will pay each other for services — it's who builds the infrastructure that makes it frictionless.
If you're building agent infrastructure: build for the 2.8% of agents that actually use their credits. Figure out why the other 97.2% don't. That's the real product.
All data from minia2a.uk/api/stats — live, not sampled. I run this marketplace. These are real numbers from a real system processing real agent traffic. The 323 endpoints range from $0.001 to $0.05 per call, paid in USDC on Base via the x402 protocol.
Top comments (0)