Agentic Vision in Gemini

One of the most immediately impactful feature groups is zooming and inspecting capabilities, which the model often employs implicitly without requiring an explicit instruction to zoom. When tasked with interpreting high-resolution building plans, for instance, Gemini 3 Flash can generate Python code that crops specific patches—such as roof edges or building sections—and reinserts them as new images for detailed analysis. PlanCheckSolver.com, an AI-powered building plan validation platform, harnessed this behavior to achieve a 5% accuracy improvement by enabling code execution with Gemini 3 Flash. The backend logs show the agentic process in action: the model iteratively extracts sub-regions, checks each against complex building codes, and visually grounds its compliance reasoning. This zoom-and-inspect workflow is invaluable for any application that deals with large, dense images, from satellite imagery analysis to medical scan interpretation, where missing a millimeter-scale anomaly could have serious repercussions.

Image annotation is another transformative feature that turns the model's internal reasoning into an explicit, shareable visual scratchpad. Instead of merely describing what it sees in text, Gemini 3 Flash can execute Python code to draw directly on the canvas, placing bounding boxes and numeric labels over objects of interest. In a demonstration within the Gemini app, the model was asked to count the number of digits on a hand; it responded by visually boxing and numbering each finger it identified, preventing the kind of off-by-one mistakes that are common in standard LLMs. This capability goes beyond simple counting—it allows the model to highlight areas of interest, mark discrepancies, or illustrate spatial relationships, making its reasoning transparent and auditable. For collaborative workflows, where a human reviewer needs to verify AI-generated conclusions, these annotated images serve as an intuitive, at-a-glance record of the model's step-by-step logic, greatly improving trust and efficiency.

Agentic Vision also excels in visual math and plotting, a feature group that tackles the well-known weakness of LLMs in multi-step quantitative reasoning. When presented with a dense table of benchmark scores, for example, the model identifies the raw data, writes Python code to normalize the values relative to a prior state-of-the-art baseline, and generates a professional Matplotlib bar chart—all without hallucinating numbers. This offloads computation from the probabilistic language model to a deterministic Python environment, so that arithmetic operations and data transformations are executed with perfect precision. The resulting visualizations are not mere approximations but exact representations grounded in the source image. Developers can use this to build dashboards that extract charts from legacy reports, automate financial statement analysis, or verify scientific data, knowing that the plotted points are generated by verifiable code execution rather than an opaque model guess.

The overall workflow of Agentic Vision is straightforward to integrate: developers enable the code execution tool when calling the Gemini API, and the model autonomously decides when and how to use Python to manipulate or analyze images. Through the Think–Act–Observe loop, it treats every image understanding task as a mini investigation, generating code to crop, rotate, annotate, or visualize, then appending the results to its context window for deeper inspection. This methodology is supported by a consistent 5–10% quality boost across most vision benchmarks when code execution is turned on. The model's behavior continues to evolve; currently, implicit zoom is robust, while actions like rotating images or initiating visual math may still need a gentle prompt nudge, but future updates aim to make these fully implicit. This agentic approach effectively gives the model a toolkit of visual manipulation primitives, enabling it to reason step-by-step and ground every answer in visual evidence.

Concrete use cases already demonstrate the breadth of this technology. PlanCheckSolver.com uses the zoom and inspect capability to validate complex building plans against code requirements, increasing accuracy and reducing manual review time. In the Gemini app, users can count objects with precision by leveraging the image annotation feature, eliminating guesswork. Data analysts can upload a screenshot of a performance table and receive a perfectly scaled bar chart, converting static reports into interactive insights. Quality control teams can inspect manufacturing images for defects, relying on the model's ability to highlight and measure anomalies. AI Studio Playground users can prototype these scenarios without writing code, simply by turning on Code Execution under Tools, making it easy to experiment with agentic visual reasoning before integrating it into production pipelines.

Agentic Vision is tailored for AI developers, machine learning engineers, computer vision specialists, and startup teams building applications that require reliable image analysis. It is currently available in Gemini 3 Flash via the Gemini API on Google AI Studio and Vertex AI, and is rolling out to the Gemini app under the Thinking model selection. While pricing details are not specified here, the capability is embedded in the Gemini model ecosystem, meaning developers pay for API usage as they would for any Gemini 3 Flash request. The target platforms span cloud, mobile, and web through standard API calls, with Python client libraries and quickstart code snippets provided in the official documentation. In summary, Agentic Vision in Gemini 3 Flash turns passive image recognition into an active, code-driven investigation, delivering the accuracy and transparency that mission-critical visual tasks demand.