Google just shared some exciting news about their “Agentic Vision” – it’s a clever new way for AI to understand images, linking what it sees with code to give answers backed by visual proof. They’re saying it boosts quality on most vision tests by a solid 5% to 10%!
<div class="media-with-label__label">
Credit: Kundra / Shutterstock </div>
</figure>
</div>
</div>
</div>
</div>So, Google has popped a neat new feature called Agentic Vision into their Gemini 3 Flash model. What’s cool about it? Well, they say it brings together visual understanding and code execution to really *anchor* AI answers in actual visual evidence. Apparently, this is a pretty big game-changer for how AI models handle images!
This Agentic Vision feature, which first dropped on January 27th, is ready for you to check out. You can access it through the Gemini API in Google’s AI Studio development tool, or via Vertex AI within the Gemini app itself.
Google explains that Agentic Vision in Gemini Flash totally transforms how AI “sees” images – it’s no longer just a quick look, but an active, intelligent process. By mashing up visual understanding with code, the model can actually *plan* to zoom in, examine, and even tweak images step-by-step. Before this, most AI models that handled both text and images would just take one static glance. If they missed something tiny, like a product’s serial number or a distant street sign, they’d often just have to guess. But now, with Agentic Vision, understanding an image becomes like a proper investigation, incorporating an active, “think, act, observe” cycle into how it processes visuals. Pretty smart, huh?
This Agentic Vision lets the AI model actually *interact* with what it’s seeing by adding annotations right onto images. So, instead of simply telling you what’s in a picture, Gemini 3 Flash can run code to draw directly on the image itself, helping to back up its reasoning. Plus, Agentic Vision is capable of understanding complex tables and even running Python code to make those findings easier to visualize. Looking ahead, Google plans to give Agentic Vision even more behind-the-scenes code-driven actions, arm Gemini models with more handy tools, and make this cool feature available in a wider range of model sizes, not just Flash.
