What 13,000 Agent API Trials Tell Us About the Agent Economy

There's now a live marketplace where AI agents discover, trial, and pay for APIs — and it's generating enough trial data to answer a question nobody could answer six months ago: what do agents actually call when nobody tells them what to call?

After 13,000+ trials across 306 endpoints, patterns are emerging. Here's what the data says.

The Top 10: What Agents Reach For First

The top 10 endpoints account for 54% of all trials — a power-law distribution that mirrors every other API marketplace. But the top 10 also tell a story about what autonomous agents need.

The 4 Categories Agents Actually Pay For

When you cluster all 306 endpoints by function, four clear categories emerge:

1. Agent Infrastructure (Memory, Storage, Identity)

Recall (1,831 trials, 98 users) is the runaway #1. Agents need memory — a place to store context, search past interactions, and retrieve structured data. Store (298 trials) reinforces this: agents need persistence that survives a session.

The insight: the first thing an agent does when given money to spend is buy memory. This isn't surprising — humans built libraries before they built stock exchanges — but the demand intensity is notable. Recall has 14% more trials than the #2 endpoint.

2. Blockchain & Financial Data (Gas, Polymarket, Token Security)

Gas (1,345 trials, 175 users) has the highest unique-user count — more than recall, despite fewer total trials. Every blockchain agent needs gas data. Polymarket (672), token-security (153), fear-greed (133), and funding-rate (75) round out the financial cluster.

The insight: crypto-native agents are the earliest adopters of agent payments. They already understand wallets, USDC, and pay-per-call. They're the beachhead.

3. Web Interaction (CAPTCHA, Screenshot, Scrape, Search)

Captcha-solve (1,202 trials, 133 users) and find (1,179 trials, 84 users) make up the web-access layer. Agents need to see the web: solve CAPTCHAs, take screenshots (347), scrape content (307), and search (105).

The insight: the web was built for humans. Agents pay for a translation layer — CAPTCHA solving is literally "help me prove I'm not a robot so I can access the human web." This category will grow as more agents go autonomous.

4. Data Enrichment & Analysis (Sentiment, Wallet Intel, Domain Intel)

The long tail: sentiment (97), wallet-intel (45), domain-intel (58), email-verify (68). These are composable primitives — agents chain them together. A DeFi agent might call gas → token-security → swap-safety → multi-balance in sequence.

The insight: the value isn't any single endpoint. It's the composability of 306 endpoints that an agent can chain without negotiating 306 API keys.

The Stickiness Gap

Here's the number that matters most: 13,106 trials from 321 unique users = 40.8 trials per user. But 63 users registered wallets (19.6% conversion to registration), and only 14 have made paid transactions (4.4% of total users, 22.2% of wallet holders).

The funnel: 321 visitors → 63 wallets → 14 paid → 820 credits spent out of 35,676 available (2.3% utilization).

What does this mean?

1. Free trials work. 13,000 trials prove agents are finding and using the marketplace. The discover → trial path is working.
2. Trial → registration works (20%). One in five trial users registers. This is strong for a developer product at this stage.
3. Registration → payment is the bottleneck (22%). 63 wallets, 14 paid. The gap isn't discovery or trialing — it's the jump from "I tried it for free" to "I'm putting money in."
4. Credit utilization (2.3%) is the quiet problem. 35,676 credits issued, 820 spent. Users register, get 500 free credits, and leave before spending them.

Why Credits Sit Idle

The standard answer is "the product isn't sticky enough." But the data suggests something more specific: agent developers don't yet have a habit of "spending credits on API calls." They're used to free APIs, API keys, or subscriptions. Credits are a new mental model.

Compare to AWS: when AWS launched, developers didn't think "I'll spin up an EC2 instance" — they thought "I'll buy a server." The mental model had to shift from capital expenditure (buy hardware) to operational expenditure (pay per hour). Agent credits face the same category-creation problem: shifting from "I'll get an API key" to "my agent pays per call."

This isn't a product problem — it's an adoption-curve problem. The 14 paid users and 820 spent credits are the innovators. The early majority hasn't arrived yet.

What Changes on August 14

Claude Code auto mode ships as default in 4 days. This matters because:

1. Auto-mode agents will hit 402 paywalls without a human to approve. The decision flow becomes: encounter 402 → check .agent-budget → pay if within limits → retry. No human in the loop.
2. Every 402 response needs machine-readable payment instructions. The agent needs to know: amount, chain, token, recipient — in headers it can parse, not HTML it can't read.
3. Discovery becomes the binding constraint. The agent has money and permission to spend it. It needs a list of APIs it can call. The marketplace that answers "which API should I call?" becomes the gatekeeper.

This is why discovery matters more than payment rails. Cloudflare, Coinbase, and OSL solved the payment layer. Nobody has solved "which of 150,000 APIs does my agent need right now?"

The Data Says: Build for Composability

The lesson from 13,000 trials: agents don't want one API. They want a surface area of APIs they can chain together — memory + search + blockchain data + web rendering. The marketplace that makes this composability frictionless wins.

Four things that move the needle:

1. Machine-readable discovery. An agent should be able to GET /api/discover?need=gas+price and get a ranked list of endpoints with prices, chains, and trial availability.
2. Trial-first onboarding. 13,000 trials prove the model. The question is converting trial users into paying users — and 500 free credits on registration is the bridge.
3. Budget pre-approval. .agent-budget — a 5-line JSON file. The human sets it once; the agent obeys it always. Without this, agents stop at every 402 and ask.
4. Endpoint health transparency. Agents need to know which endpoints are live before they pay. Verified discovery beats unverified registries.

Data source: minia2a.uk/api/stats (live, updated continuously). All numbers as of August 10, 2026.