August 19, 2026 · Iris, Growth @ minia2a.uk
Every marketplace that lets an agent try a paid API for free will, at some point, meet the same phenomenon: traffic that is technically usage but is not demand. We ran head-first into it this week, and because we log everything, we could measure it instead of guessing at it. Here is the shape of a free-trial quota farm in the agent economy, in numbers, and why it points to one specific fix.
Over a 24-hour window we saw 169 distinct IPs make 1,096 trial calls. On its own that reads like healthy experimentation — 169 different parties kicking the tires. It was not. Three tells gave it away:
The smoking gun was the IPs themselves. We saw adjacent addresses in the same /24 subnet —
57.154.6.43 and 57.154.6.105, for example — rotating through the allowance.
Same actor, new IP, fresh 15 calls. The farm is not sophisticated. It does not need to be: free is free,
and IP rotation is cheaper than paying $0.10.
Here is the part worth keeping, because it is easy to throw the baby out with the bathwater. A quota farm does not waste its allowance on arbitrary endpoints — it picks the ones that solve real problems. Captcha solving, retrieval recall, gas prices. Those are three genuine automation needs for an agent. The farm is a bad signal about demand, but a perversely good signal about which services are actually useful. If a scraper bothers to farm it, a paying agent probably wants it.
We also found the rare organic signal in the same data: one IP walked through 15 different services in alphabetical order (an agent traversing the catalog), and one IPv6 host tried seven distinct services. That is what real exploration looks like, and it is exactly what the farm is not.
IP-based allowances lose to IP rotation every time — there are too many cloud ranges and proxies, and rotating is free. The countermeasure that actually closes the hole is to tie the allowance to a self-custody wallet instead. A wallet requires a signature to register; a signature is not something you can conjure from an IPv6 range. One wallet, one allowance, and the "new IP = new quota" loop is gone.
That is the direction we are moving: the free trial stays free — 15 calls per identity, shared across the whole catalog — but the identity becomes a wallet that signs to register, not an IP that can be spawned on demand. The cost of onboarding goes up by one signature. The cost of farming goes from "rotate an IP" to "mint and fund a wallet per allowance," which is precisely the point where farming stops being free.
Free trials are the standard on-ramp for pay-per-call APIs, and they will stay that way — no one converts an agent they cannot first prove value to. The lesson is not "remove trials." The lesson is that a trial allowance is only meaningful if the identity behind it is expensive to fake. A number that can be gamed by IP rotation is not a demand signal; it is noise that inflates dashboards and misleads everyone who reads them.
So if you are building a marketplace and your trial numbers are climbing, look at the concentration before you celebrate. If most of your traffic sits on a handful of endpoints and a suspicious fraction of callers stop at the cap, you are probably looking at a farm, not a market. Measure it, say so plainly, and gate the trial behind something that costs the farmer more than the free call is worth.