Frontier multimodal models usually process an image in a single pass. If they miss a serial number on a chip or a small symbol on a building plan, they often guess. Google’s new Agentic Vision capability in Gemini 3 Flash changes this by turning image understanding into an active, tool using loop grounded in visual evidence.

Google team reports that enabling code execution with Gemini 3 Flash delivers a 5–10% quality boost across most vision benchmarks, which is a significant gain for production vision workloads.

What Agentic Vision Does?

Agentic Vision is a new capability built into Gemini 3 Flash that combines visual reasoning with Python code execution. Instead of treating vision as a fixed embedding step, the model can:

• Formulate a plan for how to inspect an image.
• Run Python that manipulates or analyzes that image.
• Re examine the transformed image before answering.

The core behavior is to treat image understanding as an active investigation rather than a frozen snapshot. This design is important for tasks that require precise reading of small text, dense tables, or complex engineering diagrams.

The Think, Act, Observe Loop

Agentic Vision introduces a structured Think, Act, Observe loop into image understanding tasks.

1. Think: Gemini 3 Flash analyzes the user query and the initial image. It then formulates a multi step plan. For example, it may decide to zoom into multiple regions, parse a table, and then compute a statistic.
2. Act: The model generates and executes Python code to manipulate or analyze images. The official examples include:
   • Cropping and zooming.
   • Rotating or annotating images.
   • Running calculations.
   • Counting bounding boxes or other detected elements.
3. Observe: The transformed images are appended to the model’s context window. The model then inspects this new data with more detailed visual context and finally produces a response to the original user query.

This actually means the model is not limited to its first view of an image. It can iteratively refine its evidence using external computation and then reason over the updated context.

Zooming and Inspecting High Resolution Plans

A key use case is automatic zooming on high resolution inputs. Gemini 3 Flash is trained to implicitly zoom when it detects fine grained details that matter to the task.

Google team highlights PlanCheckSolver.com, an AI powered building plan validation platform:

• PlanCheckSolver enables code execution with Gemini 3 Flash.
• The model generates Python code to crop and analyze patches of large architectural plans, such as roof edges or building sections.
• These cropped patches are treated as new images and appended back into the context window.
• Based on these patches, the model checks compliance with complex building codes.
• PlanCheckSolver reports a 5% accuracy improvement after enabling code execution.

This workflow is directly relevant to engineering teams working with CAD exports, structural layouts, or regulatory drawings that cannot be safely downsampled without losing detail.

Image Annotation as a Visual Scratchpad

Agentic Vision also exposes an annotation capability where Gemini 3 Flash can treat an image as a visual scratchpad.

In the example from the Gemini app:

• The user asks the model to count the digits on a hand.
• To reduce counting errors, the model executes Python that:
  • Adds bounding boxes over each detected finger.
  • Draws numeric labels on top of each digit.
• The annotated image is fed back into the context window.
• The final count is derived from this pixel aligned annotation.

Visual Math and Plotting with Deterministic Code

Large language models frequently hallucinate when performing multi step visual arithmetic or reading dense tables from screenshots. Agentic Vision addresses this by offloading computation to a deterministic Python environment.

Google’s demo in Google AI Studio shows the following workflow:

• Gemini 3 Flash parses a high density table from an image.
• It identifies the raw numeric values needed for the analysis.
• It writes Python code that:
  • Normalizes prior SOTA values to 1.0.
  • Uses Matplotlib to generate a bar chart of relative performance.
• The generated plot and normalized values are returned as part of the context, and the final answer is grounded in these computed results.

For data science teams, this creates a clear separation:

• The model handles perception and planning.
• Python handles numeric computation and plotting.

How Developers Can Use Agentic Vision Today?

Agentic Vision is available now with Gemini 3 Flash through multiple Google surfaces:

• Gemini API in Google AI Studio: Developers can try the demo application or use the AI Studio Playground. In the Playground, Agentic Vision is enabled by turning on ‘Code Execution‘ under the Tools section.
• Vertex AI: The same capability is available via the Gemini API in Vertex AI, with configuration handled through the usual model and tools settings.
• Gemini app: Agentic Vision is starting to roll out in the Gemini app. Users can access it by choosing ‘Thinking‘ from the model drop down.

Key Takeaways

• Agentic Vision turns Gemini 3 Flash into an active vision agent: Image understanding is no longer a single forward pass. The model can plan, call Python tools on images, and then re-inspect transformed images before answering.
• Think, Act, Observe loop is the core execution pattern: Gemini 3 Flash plans multi-step visual analysis, executes Python to crop, annotate, or compute on images, then observes the new visual context appended to its context window.
• Code execution yields a 5–10% gain on vision benchmarks: Enabling Python code execution with Agentic Vision provides a reported 5–10% quality boost across most vision benchmarks, with PlanCheckSolver.com seeing about a 5% accuracy improvement on building plan validation.
• Deterministic Python is used for visual math, tables, and plotting: The model parses tables from images, extracts numeric values, then uses Python and Matplotlib to normalize metrics and generate plots, reducing hallucinations in multi-step visual arithmetic and analysis.

Check out the and . Also, feel free to follow us on  and don’t forget to join our  and Subscribe to . Wait! are you on telegram? 

The post appeared first on .

 Axiom says its AI found solutions to several long-standing math problems, a sign of the technology’s steadily advancing reasoning capabilities.