India's agentic AI adoption mirrors mobile revolution: DeepMind executive

Indians embracing AI agents is similar to the mobile revolution that propelled the country ahead, said a top executive at Google’s artificial intelligence research arm DeepMind, noting that the country is playing a key role in building agentic use cases.

“We are super excited about the agentic era in India, where we are deploying multiple agents,” Seshu Ajjarapu, senior director for applied AI at Google DeepMind, told ET.

AI agents, or autonomous digital assistants capable of executing various tasks across apps for individual users, have emerged as a key narrative shaping the artificial intelligence landscape.


Speaking to ET on the sidelines of Google I/O Connect India 2026 in Bengaluru on Tuesday, Mountain View, California-based Ajjarapu said India has a prominent role in the post-training phase, where one needs to build use cases around the model, which includes agentic use cases, and scientific discovery.

This requires models to follow instructions and engage in reinforcement learning with human feedback. It involves mixing multiple languages to see how the AI model reacts and rewarding models when they do well.

“Take education as a use case. When we want to be educated, we don’t want Gemini to spit out an answer, but to teach us how to get an answer in a logical fashion,” Ajjarapu explained. “A lot of this stuff happens in India on top of the models.”

While agentic AI is a key theme, Google DeepMind’s agentic coding platform Antigravity 2.0 may have to play catch up with OpenAI’s Codex and Anthropic’s Claude Code.

Sundar Pichai, CEO of Google and Alphabet, acknowledged this in a recent podcast with the New York Times.

While the company is confident about closing this gap, Ajjarapu said coding is not Google’s only focus and they see the world in general purpose models.

“Our philosophy has always been about multimodality at the core,” he said. “Code is just one modality, audio and video are another. We try to advance all the things so that people can build on top of that.

On the rising compute cost and adoption of the company’s light-weight, open-source Gemma models, Ajjarapu said while developers may use Gemma more, for the enterprise use cases, they are seeing workload shifting to Gemini flash model, where they get the speed, and capabilities.

“I do think that people will play with multiple models,” he said. “Our belief is that we are going to offer an entire gamut of choice to the users, open and close models, not just from the Gemini model family. Even in terms of TPU stack, we’re going to offer a full stack, GPUs, TPUs, CPUs, whatever you want to use.”