Google Launches Gemini 2.0 Flash, Previews AI Agents

Google introduced Gemini 2.0 Flash on Wednesday, the first release in its new generation of artificial intelligence models. The significance lies not only in its speed and multimodal capabilities—allowing it to work with text, images, audio and video—but also in Google’s positioning of it as a foundation for agents: systems that do more than answer questions, with the ability to plan tasks and use tools.

Gemini 2.0 Flash is initially being released as an experimental version for developers through Google AI Studio and Vertex AI, Google Cloud’s enterprise platform. It is also beginning to roll out in the company’s own products, including AI-generated answers in the US search engine.

A model built to generate, not just interpret

Earlier multimodal models could already analyze a photograph, summarize a document or answer questions about a video. Gemini 2.0 Flash expands on that concept with native capabilities for generating images and audio, in addition to text. The company has shown, for example, voice conversations with expressive responses and translations that better preserve the speaker’s tone.

The model retains a context window of up to one million tokens. In simple terms, it can handle a very large amount of information within a single conversation: lengthy documents, hours of recordings or large codebases. That capability was already available in Gemini 1.5, but Google is now looking to combine it with more fluid interaction and the use of external tools.

The distinction matters because a conventional assistant receives a question and returns an answer. An agent can break a request into steps, search for data, query connected services and present a result. That process opens up useful possibilities, but it also increases the risk of errors: if a system misinterprets an instruction and acts on a tool, the consequences can go beyond an incorrect answer.

Deep Research: Research with more intermediate steps

One of the first applications of Gemini 2.0 is Deep Research, a feature for Gemini Advanced. The system can create a research plan, browse web pages and produce reports with sources.

It is not a substitute for a researcher’s work, nor does it guarantee that its conclusions will be correct. Quality depends on the sources it finds, how it interprets their contents and whether the user reviews the result. Its value lies in reducing the time needed to gather material and organize an initial synthesis, especially when comparing products, preparing a story or exploring scattered information.

Google plans to launch Deep Research in Gemini Advanced over the coming weeks.

Astra wants to understand the world through the camera

Project Astra is the assistant prototype that DeepMind first showed in May. Its goal is to enable AI to observe its surroundings through a phone’s camera, hold a voice conversation and remember what it has seen during the interaction.

Google has brought some of those capabilities to the Gemini app: real-time conversation, more languages and a more natural response when users share video. The promise is clear for everyday tasks, from locating an object to getting help while carrying out a repair. But Astra remains a project in development, not a finished product with full autonomy.

Mariner brings the agent to the browser

The most ambitious demonstration is Project Mariner, a Chrome extension prototype that observes the visible content of a web page and can move the cursor, type and click to complete tasks. Google is testing it with a small group of trusted users.

In one demonstration, Mariner could search for products and add options to a list. This is a different direction from that of a chatbot: instead of asking users to copy data between tabs, the system attempts to carry out the steps in the browser.

Even so, Google acknowledges that Mariner is a research experiment. Web agents face changing pages, ambiguous forms, payments and personal data. The challenge is not simply teaching them to press buttons, but ensuring they understand when to stop, ask for confirmation and avoid making decisions that belong to the user.

Gemini 2.0 Flash gives Google a common platform for that strategy. The race is no longer just about building models that write or converse better; it is about enabling them to operate programs and services without turning every request into a security and oversight problem.