Gemini 3.5 Flash integrates computer use

Google is integrating computer use directly into Gemini 3.5 Flash. This allows the model to see interfaces, reason about what is happening on screen, and take actions in browsers, mobile devices, and desktop computers.

Until now, this feature was only available as a standalone model based on Gemini 2.5. By adding it to the Flash version, developers can create agents that complete longer tasks and work across different applications without having to connect a separate model.

What Gemini 3.5 Flash can do

The feature is designed for agents that do more than respond to a command. They can also interact with programs and services. For example, an agent could navigate an application to check that it works correctly, update information across multiple tools, or perform repetitive tasks in a company's work environment.

Gemini could already call functions and use integrated tools such as Search and Maps. It can now combine those capabilities with visual interaction and interface control, expanding its use in:

• Continuous software testing.
• Automation of administrative and professional tasks.
• Work with business applications.
• Processes that require multiple steps and intermediate decisions.

The important difference is that the agent can operate for longer and adapt to what appears on screen, instead of being limited to generating text or returning a one-off response.

More controls for sensitive actions

Giving an agent access to a computer also introduces risks. A web page or document could include hidden instructions designed to trick it into revealing information or taking an unauthorized action. This type of attack is known as indirect prompt injection.

Google says it trained Gemini 3.5 Flash with specific adversarial examples to reduce this risk. It also offers two optional protection systems for businesses:

• Require the user's explicit confirmation before executing sensitive or irreversible actions.
• Automatically stop the task if a possible indirect prompt injection is detected.

These measures do not replace other layers of security. Google also recommends using isolated environments, human review, and strict controls over the agent's permissions.

What changes for you

For a user, this could mean assistants that complete entire processes instead of merely explaining how to do them. For example, they could review an application step by step or prepare information across multiple tools, as long as they have the necessary permissions and the system requests approval when an action is sensitive.

For a business, the benefit is the ability to automate tasks that previously required someone to copy data between applications or supervise every step. The limits will still depend on the agent's reliability, the access it receives, and its ability to detect malicious instructions.

Gemini 3.5 Flash with computer use can already be tested in a demonstration environment hosted by Browserbase. Google also offers a reference implementation and documentation for building agents with the Gemini API and Gemini Enterprise Agent Platform. What to watch now is how it performs on long, real-world tasks, where a small error can have significant consequences.