A warm-up scrolls a feed the way a person would and, now and then, likes or saves or follows — slowly, at low rates, so an account matures like one somebody actually uses rather than one that appeared overnight and started posting.
The scrolling and the tapping are the easy part. The part that matters — and the part most automation gets wrong — is knowing what is actually on the screen before you touch it.
Farmero looks before it taps
Every tap and every scroll starts by reading the screen's own accessibility tree and asking what is there. The like is not a guess at where the heart usually sits; it is resolved from this frame's rail, by the control's own identifier. If the heart cannot be found, the like does not happen — it is never a blind press at a coordinate that was right on some other screen.
That sounds like a detail. It is the whole difference, and here is the incident that proves it.
The scroll that opened an ad
An account owner reported his YouTube warm-ups "opening a lot of ads and random apps." Measured the same hour: a YouTube ad renders as a single button that covers almost the whole screen. Automation that puts its finger down at a fixed spot — the way every coordinate-based tool does — was pressing that spot with no idea what was under it. On an ad frame, that spot was the advertiser's call-to-action, and the phone was found sitting inside a full-screen trading-app ad it had tapped its way into.
The fix was not to move the fixed spot to a "safer" number. A coordinate cannot be safe on a screen nobody measured. The fix was to ask the screen where the finger can go down without pressing anything — and, if every candidate is covered by an ad, to not scroll at all on that frame rather than guess.
That is the difference between automation that looks and automation that does not, and it is not a marketing line — it is a bug that happened, and the code that came out of it.
It refuses rather than guesses
The same discipline runs through the whole warm-up: when Farmero cannot confirm what it is looking at — the feed it expects, the account it is supposed to be — it stops and says so, instead of carrying on against a screen it does not recognise. A run that refuses is a run that did no harm.
The rates stay low, per account
Engagement is deliberately sparse and capped — a like is an occasional coin-flip, saves and follows rarer still — and each account has its own steady rhythm rather than the whole fleet behaving on one average. The point of a warm-up is to look like one person's ordinary use, and one person does not like every video.
Most tools cannot do any of this, because most tools cannot see: they read back their own screenshots or fire at fixed coordinates. Farmero drives real iPhones and reads what is on them, which is why it can refuse the tap that would have opened the ad — and why the one number it reports about itself, its verified-live rate, is a measurement instead of a hope.