Gemini Agentic Vision: how it works and how to test it

What is Gemini Agentic Vision?

Google's launch announcement describes an image-analysis process that uses generated Python code to inspect evidence. Instead of relying only on the original view, the model can work with transformed images and intermediate results.

Google reported a 5–10% quality improvement across most vision benchmarks when enabling code execution with Gemini 3 Flash. That is a vendor-reported benchmark result, not an expected improvement for every customer's images.

How the think, act, observe loop works

The three stages describe what changes during an inspection:

For example, a crop may make a small label easier to inspect. But enlarging a region does not establish that a character was readable in the original. If the label remains ambiguous, "I cannot distinguish these characters" is a better result than an invented serial number.

What can it help investigate?

Google's launch examples include inspecting detailed building plans, annotating fingers while counting them, and turning numbers from a table into a plotted comparison. Those are demonstrations of the method, not proof that it is reliable on your dataset.

Google's Gemini app demonstration shows image annotation during counting. The video is edited and sped up for demonstration, as noted on screen. Source: Google's announcement.

I would evaluate these capabilities with three separate checks:

• Small details: compare the extracted text against a known label. Include images with blur or partial obstruction, not only clean samples.
• Counting: inspect the marked objects as well as the total. A correct total can hide one missed object and one duplicate.
• Visual arithmetic: check the extracted values before checking the calculation. Correct Python arithmetic cannot repair a misread input.

For a damaged-item photo, a visible crack might justify asking a follow-up question. It does not by itself establish when the damage occurred, whether the customer is eligible for a replacement, or whether an action should be taken.

How to try Gemini Agentic Vision

Start from Google's current code execution guide, not an unchanged copy of a launch snippet. The guide says image execution needs Code Execution and Thinking enabled. It recommends explicitly requesting code for tasks such as rotation or counting.

As checked September 8, 2026, the guide's image section still names Gemini 3 Flash, while its code examples use gemini-3.8-flash with the Interactions API. The model catalog separately lists gemini-3-flash-preview. Do not treat these model names or API request formats as interchangeable.

Use the example for the model and API you intend to deploy. Keep its image-input format, tool configuration, authentication setup, and response handling together. Copying only the model name into an older snippet does not verify that the request will work.

For an initial trial, Google AI Studio is a place to explore the available models and tools. Check the actual settings in your account rather than assuming that selecting a similarly named consumer-app mode exposes the same API behavior.

A small evaluation you can repeat

Build a test set with answers you can check manually. For example, use approved non-sensitive images containing ten known labels, a set of objects with an agreed count, and a table whose values you already have in text form.

Record the exact model, prompt, tool settings, and image for each run. Compare the same examples with and without the intended code-execution setup where that comparison is supported. Do not change the prompt and image quality at the same time, or you will not know what affected the result.

These are suggested evaluation criteria, not measured results from an eesel benchmark. Set your own acceptance threshold according to the consequence of an error.

Limitations that matter in production

Google documents a 30-second maximum for the code environment and says users cannot install additional libraries. It also notes variation between models and possible regressions on other tasks when code execution is enabled. Check the execution limits before designing a larger processing job.

Do not assume that enabling the tool means every answer used it. Inspect the returned execution steps when your test depends on a crop or calculation. Keep the image, intermediate result, and final answer together so a reviewer can locate the mistake.

Likewise, a successful visual test does not validate permissions in a downstream system. If the output can cause a refund, change a record, or notify a customer, test that action separately. A useful rollout starts with reviewable suggestions and a narrow set of permitted actions.

Gemini Agentic Vision pricing

As checked September 8, 2026, Google's Gemini API pricing lists these standard paid rates for the original Gemini 3 Flash Preview model:

These are Gemini Developer API rates for gemini-3-flash-preview, not a quote for every Flash model, Vertex AI, or the consumer Gemini app. Use the price for your selected model and service.

Google says enabling code execution has no separate Gemini API tool charge, but generated code and execution results contribute to billed usage. Estimate from representative requests, including failed attempts and retries, rather than pricing only the short final answer.

Test the support decision through eesel CLI

Reading a damaged-product photo and deciding what to do for the customer are different tasks. A vision model might identify a cracked casing, but a support reply also needs the order details, the replacement policy, and the authority to take action.

The eesel CLI lets you work with an eesel teammate from a terminal. It operates the same agent and workspace as the dashboard: you can inspect connected knowledge, upload a policy, run a conversation, and review activity. A person can type the commands, a script can process their output, or a coding agent such as Claude Code, Cursor, or Codex can run a setup review for you. Single results are JSON; lists use one JSON object per line.

That makes the CLI useful for testing what happens after an observation becomes a support question. It is not a command for enabling Google's Agentic Vision, and this example does not establish that eesel uses that Gemini feature.

Start with the policy and permissions

With Node.js 18.17 or newer, check the account and choose an existing test agent:

npx @eesel/cli login
npx @eesel/cli whoami
npx @eesel/cli agents
npx @eesel/cli status --agent TEST_AGENT_ID
npx @eesel/cli integrations --agent TEST_AGENT_ID
npx @eesel/cli instructions --agent TEST_AGENT_ID
npx @eesel/cli automations --agent TEST_AGENT_ID

Replace TEST_AGENT_ID with the agent you intend to use. Before chatting, review its action permissions and existing automations. Use a setup where customer-facing writes are disabled or require approval. Asking the agent not to send a message is a useful instruction, but it is not a permission boundary.

Prepare an approved, non-sensitive file named damaged-item-policy.txt containing the actual rules you want to test. Uploading it changes the selected agent's knowledge:

npx @eesel/cli files upload ./damaged-item-policy.txt --agent TEST_AGENT_ID
npx @eesel/cli files ls --agent TEST_AGENT_ID
npx @eesel/cli status --agent TEST_AGENT_ID

Check that the source is ready before evaluating an answer. Do not invent a replacement policy just to make the demonstration succeed.

Test uncertainty, not just a confident reply

Start a fresh conversation and supply a fictional case. This test uses a written observation, not an uploaded photograph, so it measures policy reasoning rather than image recognition.

npx @eesel/cli new --name "damaged-item-review" --agent TEST_AGENT_ID
npx @eesel/cli chat "Fictional test: a customer reports a cracked casing. The cause and purchase date are unknown. Use the damaged-item policy to draft the next question. Do not send messages or change records." --agent TEST_AGENT_ID
npx @eesel/cli chat "The customer now says the parcel arrived yesterday, but has not provided an order number. What is still missing before deciding on a replacement?" --agent TEST_AGENT_ID

The second chat command continues the conversation. A fresh conversation clears conversational context; it does not isolate the agent from its integrations or disable its tools. Chat can incur usage charges, so these are examples to run deliberately, not a free bulk test.

Look for an answer that separates the reported damage from facts the agent has not verified. It should request the information required by your policy, avoid inventing an order lookup, and leave the replacement undecided when the evidence is insufficient.

You can ask a coding agent to run the read-only setup checks and summarize missing prerequisites before requesting approval for uploads or chats. For example: "Review the selected eesel test agent's sources, instructions, and automations. Identify anything that could send a customer reply. Do not change settings or start a chat until I approve."

After a permitted test, inspect activity and held approvals:

npx @eesel/cli activity --agent TEST_AGENT_ID
npx @eesel/cli approvals --agent TEST_AGENT_ID
npx @eesel/cli billing

Billing is workspace information, not proof of this individual test's cost. A good terminal answer also does not prove that a helpdesk trigger, attachment reader, or customer reply works. Verify those separately in the intended channel with a controlled ticket. If your application needs Gemini's visual analysis upstream, build and test that connection separately; this CLI example does not supply it.

The dashboard above is another view of the eesel workspace, not a Gemini image-analysis result.

Choose the test that matches the job

Gemini Agentic Vision is relevant when the hard part is inspecting visual evidence. Evaluate the extracted details and intermediate results, not only the final answer's confidence.

If the next question is how a support teammate should respond under your company's policy, use eesel CLI to inspect its setup and test that decision with a fictional case. Keep visual accuracy, policy accuracy, and permission to act as separate checks.

Try eesel, then use the CLI setup guide to prepare a test agent before enabling customer-facing automation.