Agentic Vision: When AI Doesn't Just See an Image, It Investigates It

Most vision models work the same way: you give them an image, they give you one answer, done.
Google DeepMind's new "Agentic Vision" in Gemini 3 Flash breaks that pattern. Instead of a single pass, the model can actively interact with the image — zooming into a detail, cropping a region, re-examining a specific area before answering. It treats vision less like a snapshot and more like an investigation.
Think about the difference this makes for something like reading a dense chart, spotting a small defect in a product photo, or verifying a tiny detail in a document. A one-shot glance often misses that. A model that can "look closer" doesn't.
This is part of a broader shift in computer vision: from single-pass recognition toward models that reason and act on what they see, step by step. Efficient vision architectures are heading the same direction — not just processing images faster, but processing them smarter.