Gemini Is Evolving Into an AI Agent

Google DeepMind is increasingly shifting its vision for Gemini beyond a system that simply answers questions. The company is working towards AI that can understand what users want and actively help them complete tasks.

In a recent interview, Koray Kavukcuoglu, Senior Vice President and Chief AI Architect at Google DeepMind, discussed how the development of Gemini is moving towards more agent-like capabilities.

The change is significant because Google has previously indicated that agentic AI could play an important role in the future of Search.

From Answering Questions to Taking Action

Kavukcuoglu explained that Google DeepMind now sees Gemini as something closer to an AI agent than a traditional chatbot or language model.

One of the biggest influences on this development has been coding.

Working with software engineering tasks gave the team a way to explore how AI can use tools, interact with functions and work alongside people to accomplish specific goals.

Rather than focusing solely on making a model better at producing answers, the wider objective is to build AI that can take actions alongside users and on their behalf.

Kavukcuoglu suggested that coding was particularly important because software engineering involves many of the behaviours required from an effective AI agent, including planning, tool use and working through complex tasks.

He was asked about potential architectural innovations behind Gemini’s development but did not provide specific details, suggesting that some of the work taking place behind the scenes has yet to be revealed.

What Google Learned From Coding

According to Kavukcuoglu, Google’s experience developing Gemini’s coding capabilities helped the company understand more about what it takes to build useful agents.

The lessons went beyond simply getting an AI system to generate code.

The team also had to consider how an AI could function as a software engineering partner, use external tools and interact with functions that people rely on in their everyday work.

This helped Google move from thinking primarily about models towards thinking about agents and workflows.

Kavukcuoglu said the development process involved several areas of research progressing at the same time. The company could then bring those advances together as newer versions of Gemini were developed.

Some of the improvements introduced in recent Gemini versions were the result of research that had been underway for a year or longer.

AI Development Is Not Being Completely Reinvented

Despite the major changes in what AI systems can now accomplish, Kavukcuoglu pointed out that the fundamental process used to develop them has not changed as dramatically.

Google continues to rely on established approaches such as:

Deep learning
Pre-training
Reinforcement learning
Optimisation techniques

What has changed substantially is the environment in which AI operates.

Modern AI systems increasingly need to understand user intent, handle uncertainty and ambiguity, interact with tools and collaborate with people rather than simply produce a response to a prompt.

In other words, the end result can look revolutionary even when many of the underlying techniques remain familiar.

Gemini’s Agentic Capabilities Are Improving

The conversation also turned to how the AI frontier has changed.

Kavukcuoglu acknowledged that once a particular area of AI development reaches an advanced level, attention quickly moves towards the next challenge.

For Gemini, that has increasingly involved agentic coding and workflows.

Google’s work on earlier Gemini models provided the team with more insight into how people actually interact with AI agents. This included understanding how users expect an agent to behave when it is working with them on a task rather than simply answering a question.

Kavukcuoglu said this experience has given the team greater confidence in understanding what users need from an AI system acting as a partner.

The development process has therefore involved learning not only how to make the model more capable, but also how to make the interaction between people and AI more useful.

The Biggest Goal? More Intelligence

Kavukcuoglu was also asked what capability he would prioritise if he could instantly improve AI without the usual development time and resources.

His answer was relatively straightforward: greater intelligence.

He argued that more intelligent models would naturally perform a wide range of tasks better and behave in a more intuitive way.

The answer reflects a broader challenge facing AI developers. Improvements in reasoning, understanding and intelligence could potentially benefit almost every other capability that AI systems are expected to deliver.

What This Means for Google Search

The move towards agentic AI could have implications well beyond Gemini itself.

Google operates a wide range of products designed to help people complete tasks, including Gmail, Google Sheets and Maps. The company’s broader AI strategy increasingly appears to be following the same direction.

Instead of AI simply returning information, the goal is for it to understand a user’s objective and help move that objective towards completion.

For Search, this could represent an important shift.

Traditional search largely revolves around finding information and presenting relevant results. Agentic systems could eventually play a more active role by helping users carry out actions after understanding what they are trying to achieve.

For businesses and SEO professionals, this could also change how visibility in Google’s ecosystem works. Being present in search results may increasingly be only one part of the process as AI systems become more capable of understanding entities, selecting information and helping users complete tasks.

The Next Stage of Gemini

Google DeepMind’s comments suggest that Gemini is continuing to evolve from a conversational AI model into a broader system capable of interacting with tools, understanding workflows and assisting users with real-world tasks.

The technology behind AI may still rely on many established techniques, but the way those systems are being used is changing quickly.

As Google continues developing Gemini’s agentic capabilities, the distinction between an AI that answers questions and one that gets things done could become increasingly important — both for users and for the future of Search.