What 319 AI Agents Actually Pay For

August 8, 2026 — Data from minia2a.uk live marketplace

Everyone has a theory about what AI agents will pay for. Most of those theories are wrong.

We have 11,075 API trial calls from 319 unique agents across 252+ services. Here's what the data actually says.

Key finding: Agents don't pay for "AI." They pay for senses — the ability to see, read, verify, and navigate the world outside their model weights.

The Top 10: What Agents Actually Use

#	Service	Trials	Users	What it does
1	Gas Check	1,005	152	Query current gas prices across EVM chains
2	CAPTCHA Solve	1,159	133	Solve reCAPTCHA/hCaptcha challenges automatically
3	Find	987	78	Semantic search across indexed datasets
4	Recall	1,532	77	Retrieve previously stored agent memory/context
5	Web Scrape	306	61	Extract clean text from any URL
6	Token Security	151	41	Verify token contracts for honeypots, rug risks
7	Time	338	40	Current UTC time, timezone conversion, date math
8	Store	298	33	Persist key-value data with agent-keyed namespacing
9	UUID	200	32	Generate UUIDs (deterministic and random)
10	Polymarket	456	49	Query prediction market odds and volumes

What This Tells Us

1. Infrastructure > Intelligence

The top services are not AI models. They're utility functions that agents need to operate in the real world: check gas before sending a transaction, solve a CAPTCHA to access a website, look up the current time, generate a unique ID.

This matches the architectural reality: LLMs are good at reasoning but blind to the outside world. The first layer agents need isn't more intelligence — it's sensory input.

2. On-chain context is the #1 demand category

Gas checks (152 users), token security (41 users), and Polymarket data (49 users) together form the dominant demand cluster. Agents operating on-chain need real-time context that their training data can't provide: current gas prices, token safety scores, live market odds.

The message from the data: if you're building for on-chain agents, start with gas + token safety APIs.

3. CAPTCHA solving is a universal pain point

133 unique agents — more than any service except gas — have tried CAPTCHA solving. This isn't surprising when you think about it: an agent that can browse the web but can't solve a CAPTCHA is an agent that's locked out of half the internet.

133 agents means 133 different use cases where the flow was: navigate → hit CAPTCHA → call solver → continue. This is agent infrastructure, not a niche.

4. Memory is unsolved

Recall (1,532 trials) and Store (298 trials) together represent the agent memory stack: store data, retrieve it later. The fact that recall has the highest trial count of any service (1,532) suggests agents are actively wrestling with the context window problem.

Every agent developer eventually hits the same wall: "I need this agent to remember something across sessions." The demand data confirms it.

The Funnel: From Trial to Transaction

Stage	Count	Conversion
Trial users	319	—
Wallet registered	56	17.5%
Made a payment	14	25% of wallets
Total volume	$12.75	14 transactions

Honest assessment: 11K trials → 14 paid transactions. The trial experience works. The "first payment" moment is the bottleneck. We're in the very early innings of machine-to-machine payments — agents are exploring, testing, and kicking the tires. The habit of paying for API calls hasn't formed yet.

What This Means for Builders

If you're building APIs for agents, the demand data points to three rules:

Make it a sensor, not a brain. Agents already have reasoning. They need eyes, ears, and hands. Gas prices, CAPTCHA solving, token verification — these are sensory inputs, not intelligence services.
Price for micro-calls. The median service here costs fractions of a cent per call. Agents make dozens or hundreds of calls per task. Your pricing needs to work at that granularity.
Support x402 from day one. If your API requires a human to sign up, get an API key, configure billing, and monitor usage — you've lost the agent customer. Agents need protocol-level discovery and payment. x402 provides both.

Bottom line: The AI agent economy is real but early. 319 agents are actively testing paid APIs. They're not paying much yet ($12.75 total), but they're voting with their trials — and they're voting for infrastructure, not intelligence. The opportunity for builders is clear: give agents the senses they're missing.

Data source: minia2a.uk/api/stats, August 8, 2026. All numbers are live and verifiable. This analysis reflects real agent behavior on a production x402 marketplace.