What 14,856 Agent Trial Calls Tell Us About Auto Mode Spending

In 4 days, Claude Code auto mode becomes the default. A million-plus agents will have the ability to make HTTP requests, encounter 402 Payment Required responses, and decide — autonomously — whether to pay.

Most of the conversation has been about infrastructure: payment rails, protocol headers, wallet formats. That matters. But infrastructure tells you what can happen. It doesn't tell you what agents will do when they have money and permission.

We've been running a live x402 marketplace for agent-paid APIs. Here's what 14,856 real trial calls from 322 unique agents across 247 paid endpoints tell us about how autonomous agents actually behave.

1. Agent API demand follows a brutal power law

The top 10 endpoints account for 54% of all trial calls. The top 3 — gas price lookup, CAPTCHA solving, and crypto data search — alone represent 33%.

Below the top 10, usage drops fast. The 100th-most-popular endpoint has 26 trials. The 200th has 7. Half the endpoints have under 10 trials total.

What this means for auto mode: When agents have spending power, they won't distribute it evenly across 306 endpoints. They'll concentrate on 10–20 high-utility calls. The long tail of "nice to have" endpoints will go unused — not because they're bad, but because agent workflows optimize for the minimum set of external dependencies.

For API providers: being in the top 20 is the difference between 1,000+ calls and single digits. The ranking is driven by utility frequency, not pricing. Gas price gets called 200× more than email verification because agents check gas on every transaction but verify email once per session.

2. The trial-to-wallet conversion is 19% — and that's normal

322 unique agents have used free trials. 63 have registered wallets. That's 19.6%.

In human SaaS, a 19% free-to-registered conversion would be exceptional (industry average: 2–5%). But for agents, the benchmark doesn't exist yet — this is one of the first datasets.

The more interesting number is downstream: of 63 registered wallets, 52 have 500+ credits and have never made a single API call. They registered, received credits, and didn't come back.

What this means for auto mode: Registration is not activation. An agent that creates a wallet is not the same as an agent that spends from it. The gap between "has credits" and "uses credits" is where auto mode changes the game — because in auto mode, the agent doesn't need the human to approve each spend. The friction moves from "ask permission" to "stay within budget."

Our hypothesis: when auto mode removes the permission prompt from the loop, the 52 dormant wallets will start spending — because the human bottleneck is removed. We'll know in 4 days.

3. Trial exhaustion is the conversion trigger

The most common path to registration: agent hits the 15-trial limit on a high-frequency endpoint (gas, CAPTCHA), receives a 402 with x-402-amount: 0 and x-402-register: POST /api/v1/register-simple, and the agent (or its human) follows the registration path.

This tells us something important about the 402 response design: the x-402-amount: 0 header is critical. It tells the agent "this is not a payment demand — this is a registration invitation." Without it, the agent treats trial exhaustion the same as credit exhaustion, and stops.

What this means for auto mode: Every API with a trial tier needs two distinct 402 responses:

# Trial exhausted (free registration)
HTTP/2 402
x-402-amount: 0
x-402-register: POST /api/v1/register-simple {"name":"..."}

# Credits exhausted (payment required)
HTTP/2 402
x-402-amount: 100
x-402-chain: base
x-402-token: USDC
x-402-recipient: 0x...
x-credits-required: 1

If your 402 doesn't distinguish these two states, the agent stops at the first one and never reaches the second.

4. Agent workflows cluster around 3 jobs

Looking at which endpoints get called together (same IP, same hour):

1. Transaction pipeline: gas → chain-info → tx-decode → event-decode. Agents preparing and verifying on-chain transactions.
2. Research pipeline: web-search → web-scrape → summarize → sentiment. Agents gathering and synthesizing information.
3. Validation pipeline: token-security → swap-safety → address-parse → abi-lookup. Agents checking safety before interacting with contracts.

Each cluster is 3–5 endpoints chained together. An agent on the transaction pipeline calls gas price on every block, not once per session. An agent on the research pipeline calls web-scrape once per query, not per block.

What this means for auto mode: Pricing should follow access pattern, not endpoint type. A per-block endpoint (gas) at $0.01/call costs an agent $0.05/minute in a high-activity period — fine with a $5/day budget. A per-query endpoint (web-scrape) at the same price might cost $0.50/research task — also fine. The key is predictability: the agent needs to know cost-per-call before it builds the pipeline, which is why x-credits-required belongs on every response.

5. The dead endpoint problem isn't theoretical — it's 25%

An independent scan of our 306 services found ~25% return 404 or are inactive. These endpoints still appear in directory listings but don't respond. An auto-mode agent that tries 4 endpoints and hits 1 dead one isn't just wasting a call — it's learning that "this directory is unreliable," and may avoid the entire platform.

What this means for auto mode: Discovery platforms need automated health verification with auto-deactivation. Listing an endpoint that doesn't respond is worse than not listing it — it poisons the directory's reliability score in the agent's learned behavior.

6. What we expect to change on August 14

Three predictions backed by the data:

1. Trial volume 3–5× increase. Auto-mode agents will probe endpoints at higher frequency because there's no human in the loop to say "stop testing." The 14,856 number will grow fast.
2. Wallet activation from dormant credits. The 52 wallets with unused credits will start spending — not because the agents suddenly want different things, but because auto mode removes the "ask human" step that currently blocks every spend.
3. Power law intensifies. As more agents enter, the top 10 endpoints will pull further ahead — because new agents discover endpoints by watching what other agents use, creating a reinforcement loop.

What API providers should do now

Four things, each under 15 minutes:

1. Add 6 payment headers to every 402 response. x-402-amount, x-402-chain, x-402-token, x-402-recipient, x-credits-required, x-402-register (when trial-exhausted). Without these, agents can't pay even if they want to.
2. Distinguish trial-exhausted from credits-exhausted. x-402-amount: 0 = registration path. x-402-amount: >0 = payment path. Different headers, different agent behavior.
3. Test your endpoint. curl -s https://minia2a.uk/auto-mode-check | bash -s -- https://your-api.com/endpoint — 9 signals, pass/fail in 2 seconds.
4. List on a discovery platform with health probes. Being listed isn't enough. Being listed and verified working is what auto-mode agents will select for.

All data from minia2a.uk/api/stats, measured August 10, 2026. We'll publish a follow-up analysis on August 17 comparing pre- and post-auto-mode behavior.