15,136 Agent API Trials Reveal What Autonomous Agents Actually Want to Buy

#ai #webdev #programming #crypto

The Data

Claude Code auto mode goes default on August 14. When it does, agents will have permission to spend autonomously for the first time. The question is: what will they actually buy?

We have 15,136 data points that suggest an answer.

Every free trial on minia2a is an agent making a real API call — not a signup, not a pageview. Categorize all 252 trial-enabled endpoints by what agents actually call, and the pattern is clear:

Category	Trials	Share	Top Endpoint
Crypto/DeFi	3,648	24.1%	Multi-Chain Gas Price (226 users)
AI/ML	3,618	23.9%	Mythos x402 Recall (122 users)
Utility	1,958	12.9%	Time Utilities (102 users)
Security	1,272	8.4%	CAPTCHA Solver (133 users)
Web/Data	1,111	7.3%	Web Scraper (64 users)
Finance	413	2.7%	Trading Signals (20 users)
Other	3,116	20.6%	—

The Top 10 by Revealed Preference

Revealed preference — what agents call, not what they say they want:

Mythos x402 Recall — 1,997 trials, 122 users | AI/ML agent memory retrieval
Multi-Chain Gas Price — 1,685 trials, 226 users | Broadest reach of any endpoint
Mythos x402 Bazaar Find — 1,294 trials, 101 users | Agent service discovery
CAPTCHA Solver — 1,203 trials, 133 users | $0.001/call, the entry drug
Time Utilities — 970 trials, 102 users | Universal agent dependency
Polymarket Odds Live — 897 trials, 109 users | Prediction market data
URL Screenshot — 381 trials, 24 users | $0.01/call, high intent
UUID Generator — 348 trials, 41 users | Utility, low price
Web Scraper — 310 trials, 64 users | Data extraction
Mythos x402 Memory Store — 300 trials, 34 users | Agent persistence

Three Patterns in the Data

1. Gas prices are the universal dependency. 226 unique agents queried gas prices — more than any other endpoint. Every agent touching a blockchain needs this. It is the "hello world" of agent-paid APIs.

2. CAPTCHA solving is the entry drug. At $0.001/call (the cheapest paid endpoint), it has 133 users — third-highest trial count. Agents hit CAPTCHA walls, discover they can pay $0.001 to bypass them, and learn the payment flow on the cheapest possible endpoint.

3. Crypto + AI = 48% of all demand. These two categories alone account for nearly half of all trial calls. This is the core of the agent economy right now: agents need on-chain data (gas, prices, odds) and AI infrastructure (recall, find, store).

The 0.22% Credit Utilization Rate

Here is the number that looks alarming until you understand the context:

379,900 credits issued (752 agents × 500 credits each)
842 credits spent
0.22% utilization

This is not a product failure. It is the pre-auto-mode baseline.

Today, a registered agent cannot spend autonomously. Every paid API call requires a human operator to manually authorize the payment. The 14 paid transactions on the platform all came from developers who manually triggered payments.

August 14 changes this. When Claude Code auto mode is the default, agents can autonomously decide: "I need the current gas price → that endpoint costs $0.005 → it is within my $5/day budget → pay and proceed."

Three things unlock simultaneously:

Auto mode default — agents execute without human approval, gated by a classifier that catches 89% of dangerous actions (vs 13.6% for fatigued humans)
Classifier tokens are free — Anthropic announced Aug 10 that safety classifier costs are no longer billed
.agent-budget standard — {daily_limit_usdc: 5, max_per_call_usdc: 1} gives agents a hard spending cap with machine-readable 402 headers

What This Means for API Builders

If you are building an API that agents might pay for:

Crypto/DeFi and AI/ML are the proven categories. 48% of all trial demand is here.
Price below $0.01. 7 of the top 10 endpoints cost $0.005 or less. Agents optimize for cost.
Add machine-readable 402 headers. Without x-402-amount, x-402-chain, x-402-token, and x-402-recipient, an auto-mode agent cannot pay you. Period.
Offer trials. 15,136 trials happened because there was zero friction — no wallet, no signup, just curl.