The Gemini 3.5 flash drive has built-in computer functionality

Google has integrated computer manipulation capabilities into its speedy and efficient Gemini 3.5 Flash model, which developers use to complete various text, image and function call generation tasks. With this integration, Gemini 3.5 Flash can interact with your computer by taking over your browser, mobile apps and desktop programs. Previously, computer manipulation was only available in the Gemini 2.5 model. But now it’s part of Flash, so you don’t need to change models to use it.

Multi-step processes that happen over time are where a computer manipulation AI really shines. Instead of simply clicking buttons, a computer manipulation AI should be able to complete complex flows spanning several different programs. In its announcement, Google demonstrates how Gemini 3.5 Flash can complete these types of flows — such as launching a browser, entering information into a form, going to a spreadsheet, looking at results and repeating the process. This opens up opportunities for enterprise automation use cases that typically require robotic process automation solutions.

1Already running in production

Browserbase, Browser Use and UiPath have all integrated this technology into their products. Examples of how this can be used include automated software testing (by automatically clicking around in your app like a human would), automating knowledge work (by reading dynamic websites instead of static ones) and handling documents in enterprise environments. Rather than being a theoretical idea demonstrated in a research demo, this is a product that companies can start paying for and using right now.

2How Google handles the safety problem

With the ability to manipulate computers comes the possibility of malicious websites or documents tricking an agent into performing actions it shouldn’t. To prevent this from happening, Google has built in several security measures. First, the model has been trained using adversarial learning to detect and prevent injection attacks. If an attack is detected, the task will be immediately aborted without completing. Additionally, for critical operations such as form submissions or message sends, users can be prompted to confirm that they want to continue.

To further secure their systems, Google suggests implementing extra precautions like sandboxing the agent so it doesn’t have full access to the underlying system, having humans involved in important decisions and setting strict permission policies on what the agent is permitted to do. While these features aren’t implemented out of the box, they are things that developers should consider adding to their system.

3Why this matters for builders

Computer manipulation has always been the final piece of the puzzle in AI workflows. You could generate text, process data or even code, but if you wanted an AI to open a browser and do something with the output, you had to integrate some type of automation tool. With Gemini 3.5 Flash, this functionality is now included in a model that many developers may already be familiar with. This means that customer service agents can automatically look up information about customers in your CRM, research assistants can read dynamic websites and quality assurance tools can test your user interface by clicking around just like a human would.

You can access this functionality using the Gemini API or the Gemini Enterprise Agent Platform. Browserbase also offers a demo environment where you can play around with it before integrating it into your own products.