What 300 Agent Services Reveal About What AI Agents Actually Need

The Dataset

Let's look at real usage data from a marketplace with 323 x402 pay-per-call services, organized by the categories agents actually use. The top 20 endpoints by trial call volume reveal patterns that should shape how we think about the agent economy.

Category 1: Web Interaction > Everything Else

This is the clearest signal in the data: the #1 thing agents need is access to the web. Not APIs. Not structured data feeds. The open web — with all its messiness, CAPTCHAs, JavaScript-rendered content, and anti-bot protections.

The most striking stat here is CAPTCHA solving: 133 unique users, the highest of any endpoint. Every one of those users is an agent that hit a CAPTCHA wall, couldn't proceed, and needed help. This isn't a niche — it's the primary interaction failure mode for autonomous agents on the web.

Category 2: Blockchain Infrastructure

Gas data has 146 users — the most of any endpoint. Every DeFi agent, every trading bot, every MEV searcher needs to know what gas costs right now. This isn't a "nice to have." It's table stakes for any onchain agent.

But look at the pattern: security checks outnumber pure data queries. Token security (151 calls, 41 users) and swap safety (97 calls, 18 users) together represent agents doing due diligence before executing transactions. Agents aren't just consuming blockchain data — they're verifying safety before acting. This is the accountability layer in practice: the agent says "I'm about to swap this token, let me check if it's a honeypot first."

Category 3: Prediction Markets & Intelligence

Polymarket data at 443 calls and 49 users tells you that agents are actively trading on prediction markets. This is a real, measurable use case. Agents are pulling market data, computing probabilities, and executing trades — all programmatically.

The Fear & Greed index at 125 calls with only 5 users means those 5 users are heavy consumers — averaging 25 calls each. This is the pattern of a production agent, not a curious developer testing an endpoint.

Category 4: Format Conversion & Utility

This category surprises people. UUID generation? Random numbers? Base64? These are trivial functions that any agent could do locally. Why are agents paying for them?

The answer: agents run in constrained environments where not all libraries are available. A browser-based agent can't import uuid. A sandboxed agent doesn't have base64 in its runtime. These "trivial" endpoints fill gaps in the agent's execution environment — and agents are willing to pay for the privilege.

It's also about consistency and auditability. An agent that generates UUIDs via an external service gets the same format every time, with a traceable source. An agent that uses a local library might get different UUID versions across environments.

The Patterns That Matter

1. Agents Do Boring Work

The most-used endpoints are infrastructure utilities, not flashy AI features. CAPTCHA solving. Gas data. UUID generation. Web scraping. These are the digital equivalent of plumbing and electrical — not what futurists imagined the "agent economy" would look like, but exactly what agents need to get real work done.

2. User Count ≠ Call Volume

Some endpoints have few users but massive per-user volume (Fear & Greed: 5 users, 25 calls each). Others have many users but low per-user volume (CAPTCHA: 133 users, 8.4 calls each). The first pattern indicates production agents in automated loops. The second indicates developer exploration.

This matters for API pricing: production agents need volume discounts. Exploratory developers need generous free tiers. One pricing model can't serve both.

3. Security Is a Horizontal Need

Token security checks. Swap safety. Domain intelligence. Wallet intel. These aren't one category — they're a security layer that sits across every other category. Agents don't just need data; they need verified, trustworthy data with safety signals attached. This is where the accountability layer and the utility layer converge.

What This Means for the Agent Economy

The demand is real. 10,000+ trial calls from 300+ unique agents is not a rounding error — it's a signal. But the signal isn't pointing toward the AI-app-store future that a16z pitch decks describe. It's pointing toward something more like AWS for agents: a utility layer of operational services that agents can call on demand, pay for in real time, and compose into workflows.

The agents that need CAPTCHA solving today will need identity verification tomorrow. The agents that check token security today will need full smart-contract audits tomorrow. The agents that track gas prices today will need MEV-aware transaction routing tomorrow. The demand ladder is clear — if we build the rungs.

All usage data is from a live x402 marketplace with 323 services and 10,000+ trial calls. Categories are based on actual call volume, not marketing categories. The patterns are real.