Google Ships Agentic Vision in Gemini 3 Flash, Letting the Model Write and Run Its Own Image Analysis Code
Google's new Agentic Vision feature lets Gemini 3 Flash autonomously zoom, crop, annotate, and analyze images by generating and executing Python code — delivering a 5-10% quality boost across vision benchmarks.
Google released Agentic Vision as a new capability in Gemini 3 Flash, and the approach is architecturally interesting: rather than relying solely on the model's visual encoder to interpret an image in a single pass, Agentic Vision allows the model to write and execute Python code that manipulates the image — zooming into regions of interest, cropping, annotating, and re-examining — before producing a final answer. The feature was announced by @GoogleAI and detailed by Google developer advocate @_philschmid, who described it as combining "visual reasoning with code execution."
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.