Unlocking a New Era of Visual Intelligence: Meet Agentic Vision in Gemini 3 Flash
January 27, 2026
Imagine a revolutionary advancement in artificial intelligence that transforms how machines understand and interact with the visual world. This is precisely what Agentic Vision, a groundbreaking feature of Gemini 3 Flash, aims to achieve by seamlessly integrating visual reasoning with code execution, allowing for responses that are firmly rooted in visual evidence.
Rohan Doshi
Product Manager, Google DeepMind
Typically, frontier AI models like Gemini analyze the environment in a singular, static moment. When they overlook subtle details—such as a serial number on a microchip or a far-off street sign—they often have to resort to guesswork. But with the introduction of Agentic Vision within Gemini 3 Flash, this static approach evolves into an active, investigative process. This innovative capability transforms image comprehension into a dynamic exploration, enabling the model to develop strategies for zooming in, inspecting, and manipulating images step-by-step, all while grounding its conclusions in visual data.
The addition of code execution in Gemini 3 Flash delivers a remarkable improvement in quality, consistently enhancing performance by 5-10% across various vision benchmarks.
Agentic Vision: A Leap Forward in AI Capabilities
Agentic Vision incorporates a proactive "Think, Act, Observe" cycle specifically designed for tackling image understanding tasks:
- Think: The model begins by carefully analyzing the user's query along with the initial image, crafting a multi-step plan for its next actions.
- Act: It then generates and executes Python code to actively manipulate images—whether that's cropping, rotating, or annotating—or conducting analyses such as running calculations or counting objects.
- Observe: The modified image is then integrated into the model's context window, enabling it to examine this new information with improved context before producing a final response.
Real-World Applications of Agentic Vision
By enabling code execution through the API, a wealth of new behaviors can be unlocked, many of which are showcased in our demo application available here: Google AI Studio Demo. From major applications like the Gemini app to emerging startups, developers have begun to integrate this capability, uncovering numerous use cases, including:
Zooming and Inspecting
Gemini 3 Flash is adept at automatically zooming in when it detects intricate details. For instance, PlanCheckSolver.com, an AI-driven platform for validating building plans, achieved a 5% accuracy boost by utilizing code execution with Gemini 3 Flash to methodically inspect high-resolution inputs. The backend logs demonstrate this agentic process, where Gemini 3 Flash generates Python scripts to crop and analyze specific sections (like roof edges) as new images. By reintegrating these cropped sections into its context, the model visually grounds its reasoning, ensuring compliance with complex building regulations.Image Annotation
With Agentic Vision, the model can engage with its environment more interactively by annotating images. Rather than merely describing what it sees, Gemini 3 Flash can execute code to draw directly on the canvas, grounding its reasoning in real-time. In one example, when tasked with counting the fingers on a hand within the Gemini app, the model employs Python to create bounding boxes and numeric labels over each identified finger. This "visual scratchpad" guarantees that its final count is based on an exact understanding of the image.Visual Math and Plotting
Agentic Vision is capable of interpreting dense tables and executing Python code to visualize the results effectively. Standard large language models (LLMs) tend to make errors during complex visual arithmetic due to hallucination effects. However, Gemini 3 Flash circumvents this issue by delegating computation to a reliable Python environment. In a demonstration from our app, the model extracts raw data, writes code to normalize previous state-of-the-art metrics to a baseline of 1.0, and produces a professional-looking bar chart using Matplotlib. This process replaces uncertain guessing with verifiable computational execution.
Looking Ahead
We are just scratching the surface with Agentic Vision.
- More Implicit Code-Driven Behaviors: Currently, Gemini 3 Flash excels at making implicit decisions about when to zoom in on minute details. While other features, such as rotating images or performing visual math, require explicit prompts, we are working diligently to make these behaviors automatic in future updates.
- Expanded Toolset: We are also exploring how to enhance Gemini models with additional tools, including web searches and reverse image searches, to further solidify their understanding of the world.
- Wider Model Range: Furthermore, we plan to extend this capability to other model sizes beyond just Flash.
How to Get Started
Agentic Vision is currently available via the Gemini API in Google AI Studio and Vertex AI. It will soon be rolling out in the Gemini app (accessible by selecting "Thinking" from the model dropdown). Developers are encouraged to explore the demo in Google AI Studio or experiment with this feature in the AI Studio Playground by activating "Code Execution" under Tools. For a deeper dive, check out our developer documentation.
In conclusion, the introduction of Agentic Vision represents a significant innovation in visual reasoning and interaction within AI. As we continue to enhance this technology, we invite developers and users alike to share their thoughts: What applications do you see for such capabilities? Are there aspects of this approach that you find particularly exciting or concerning? Your feedback is invaluable as we navigate this evolving landscape together.
