What 14,000 Agent API Calls Reveal About the Agent Economy

In 4 days, Claude Code auto mode goes default. Millions of autonomous agents will have spending authority and no human in the loop.

What will they buy?

I analyzed 14,223 real agent-to-agent API trials across 306 pay-per-call endpoints. The data reveals a pattern that should change how you think about building for the agent economy.

The Top 10: Infrastructure Dominates

Endpoint	Trials	Users	Category
Search Memory	1,903	101	Infrastructure
Multi-Chain Gas Price	1,487	188	Infrastructure
Service Discovery	1,229	85	Infrastructure
CAPTCHA Solving	1,202	133	Infrastructure
Timestamp/Clock	821	69	Infrastructure
Polymarket Odds	776	77	Data
Screenshot	376	24	Infrastructure
UUID Generation	316	36	Infrastructure
Web Scrape	307	62	Infrastructure
Key-Value Store	298	33	Infrastructure

The top 4 endpoints alone account for 41% of all trials. Every single one is infrastructure.

AI Wrappers: Ghost Town

AI Tweet Thread: 6 trials, 1 user. AI README: 7 trials, 2 users. AI Tagline: 6 trials, 1 user. 15 AI content endpoints collectively get less traffic than a single UUID generator.

Why

1. Agents already have an LLM — they do not pay for what they can generate themselves
2. Agents cannot generate external data — gas prices, screenshots, CAPTCHA solutions require external systems
3. Infrastructure is composable — memory + scraping + gas = a research agent
4. Determinism commands a premium — autonomous agents with budgets need predictable costs

Discovery Itself Is the #3 Endpoint

Agents actively search for tools to use. But 76% of registered x402 endpoints are dead. The marketplace that solves verified discovery captures the first wave of autonomous agent traffic on Aug 14.

When millions of agents gain spending authority overnight, they will not pay for your GPT wrapper. They will pay for infrastructure primitives.

Full analysis with data: https://minia2a.uk/blog/agent-economy-trial-data-august-2026