Aller au contenu

Blog

Automatisation visuelle agentique IA vs déterministe en 2026 : la vérité factuelle

OculiX 3.0.4 face à AskUI en 2026 — décortication factuelle de l’automatisation visuelle agentique IA contre déterministe. Sept affirmations trompeuses corrigées, plus l’arithmétique tarifaire qui explique pourquoi la facturation à l’inférence casse à l’échelle QA.

The engine assumed one screen. The job needed four.

I cut a 300-test visual suite from seven hours to under two — on an 8 GB VM that used to fall asleep mid-run. The fast part is the headline. The hard part was that the engine, like almost every visual-automation tool, was built for a single screen. Here’s how I rebuilt isolation on top of it — and how we’re going to make it native.

Why automation tools cannot type Chinese, and what OculiX does instead

An OculiX user filed a bug. Typing two Chinese characters into an input field silently produced garbage. The fix took fifteen lines of Java, but the explanation reaches back to how AWT was designed thirty years ago around the assumption that one character equals one keystroke. Here is what happens when that assumption meets the rest of the world.

Storing spatial memory in a PNG: how we made visual UI matching 6.5× faster in OculiX

A weekend afternoon of measurement turned into a structural optimization for OculiX: persistent locator coordinates embedded in the PNG file itself, surviving git clone and cold-start JVMs. Measured ×6.5 on the find. Here is what we found, what we built, and what it changes for visual automation suites in CI.