Can a factory of AI agents launch products — and kill them — on real revenue signals?

Hypothesis

With frontier models, building a product is cheap. The moat moved to the kill filter: deciding fast, on evidence, which of a hundred ideas deserves to live. So instead of building one startup, build the machine that runs many small market tests in parallel — with pre-committed kill thresholds a human can't sweet-talk.

What we built

An AI factory that runs biweekly cohorts: ~10 ideas enter, agents harvest pain-signals from real communities, and at most 2 ideas graduate to a 14-day market test with a live checkout from day one. The gates are written down before the test starts — minimum signups, activation rate, absolute week-two returners, at least one payment. Miss the gate, the experiment dies, publicly. Winners graduate to their own domain and a structured launch playbook.

Experiments will live on their own isolated subdomains, with a public gallery — including the kill log. This notebook is the narrative layer; the factory floor is the machine.

Learnings

• Pre-committed thresholds are the whole point. Every founder moves their own goalposts; a factory with written gates cannot.
• The first cohort's harvest already taught us something uncomfortable: the pain-signals we expected in one vertical simply weren't in the corpus. A cheap, fast "no" is the product working as designed.
• A public kill log is calmer and more credible than a wall of launches.

Log

• 2026-07-11 — Factory spine live on our server: harvest, digest, and edge-scan timers running, GDPR scrub gate in place, template + gallery repos built. First cohort harvesting.