Jev Daily

ILIAS ISM researched 6,823 gyms with a 100-agent swarm for $200

One prompt, 97 countries, 2,864 drop-in prices, and no accuracy sample attached to any of it.

ILIAS ISM rebuilt a site into a global CrossFit and HYROX atlas off what he calls one supermegaprompt, using a swarm of more than 100 agents to research 6,823 gyms across 97 countries. He puts the total cost at about $200.

ILIAS ISM
@illyism
X
A 100+ AI agent swarm deep-researched 6,823 gyms across 97 countries
Nov 12, 2026 · View on X

The output he lists is 4,279 city guides, 2,864 real drop-in prices, 10,759 website screenshots taken, and every claim linked to its source. That last item is the part worth looking at. Price data for gyms is scattered across booking pages, Instagram bios and PDFs, and it goes stale, so a pipeline that attaches a source URL to each number at least makes the errors findable rather than invisible.

What the post does not say

There is no accuracy sample. Nobody has checked a random hundred of the 2,864 drop-in prices against the gyms' own pages, including, as far as the post goes, ILIAS ISM. So the honest reading is that a swarm produced 2,864 price claims for $200, not that it produced 2,864 correct prices for $200. Those are different products, and the second one is the one an actual user of a gym atlas needs.

The $200 is also a single build on a single run. It is a cost for one pass over 6,823 gyms, not a steady state figure for keeping prices current, and the post does not say how often the data is refreshed or what a refresh costs. Screenshots are the cheap part. Re-verifying prices is the part that recurs.

For people building on this, the number to take away is roughly three cents of compute per gym researched, and the number nobody has yet is the hit rate. If you are tempted to copy the pattern for your own directory, budget a manual audit of a sample before you publish the numbers as facts, because the cost of being wrong on a price lands on your domain, not on the swarm's.

ILIAS ISM has posted agent pipelines here before, including an OSINT run he used to file takedowns against a clone of his product.

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.