Google AI Mode, DeepSeek V3.1 deliver new agentic features

Agentic capabilities in AI Mode

Google added agentic AI capabilities within AI Mode. The search tool can now browse across multiple reservation platforms and websites, presenting users with a list of restaurants and links to the booking pages.

Google said AI Mode does this by using Project Mariner's browser agent capabilities and the power of Knowledge Graph and Google Maps to help users take actions on the web.

The cloud provider is partnering with various organizations, including OpenTable, Ticketmaster, StubHub and Booksy, to provide the experience.

Google said that those using the AI Mode experiment in Labs will also start seeing results tailored to personal preferences and interests, beginning  with dining-related topics.

AI Mode also now has a new link-sharing capability, allowing users to share AI mode responses. It is available in the U.S.

The new AI Mode agentic capability comes one month after Google introduced a business calling capability in search.

"This is something Google has been working on for a while," said Bradley Shimmin, an analyst at Futurum Group. He said Google showcased several years ago at its IO event the ability to use natural language to book a reservation at a restaurant.

The agentic capability within AI Mode and the different agents Google has released in the past few months, such as the BigQuery Data Agent, Notebook Agent, GitHub Agent, Looker code Assistant and Database migration agent, are working toward the same objectives, he said.

"Either building tooling for agents, meaning AI, or building tooling for humans that use AI, and whether you're just trying to talk to your data or book a restaurant reservation, the same technologies are at play here," Shimmin said.

What's also interesting about the agentic capabilities in AI Mode is the companies Google partnered with.

"By bringing together a partner ecosystem, Google can now enable agentic workflows for automated actions and deliver highly personalized experiences to its users," said Arun Chandrasekaran, an analyst at Gartner.

He added that concerns around privacy, security and oversharing with AI technology are significant and must be addressed. Addressing these challenges is especially important because the upgrades to these tools have the potential for misuse.

"These tools [are] enabling not just data professionals but also script kiddies and state actors alike to automate and scale attacks on businesses and persons," Shimmin said.

For example, if a user gives an AI model access to their email archive or phone book and somebody sends an email with a prompt injection attack that the model reads, he added, that could lead to a lot of risk.

"This is an area that right now the industry is running a little too fast toward … but I do think there are ways to mitigate those concerns," he said. "We have the tools in hand to do it. We need to prioritize doing so."

Despite the privacy issues, consumers trust Google and that remains its superpower, said Nikhil Lai, an analyst at Forrester.

The big challenge with the new agentic capability might be how it affects Google's partners. If consumers make reservations through AI Mode as opposed to going to OpenTable or, if Google extends the capability to order concert tickets, that will affect the traffic for those publishers, Lai said.

"Google has to find a way to partner with publishers without stifling traffic from them and without undermining their relationship with publishers who they need for information," he said.

DeepSeek V3.1

Just as Google is upgrading AI Mode, DeepSeek upgraded the foundation model powering its AI chatbot.

DeepSeek V3.1 is a hybrid reasoning model, which means it has a thinking and non-thinking mode. The model also has a longer context window. According to the vendor's WeChat account, the model was also designed for a home-grown chip that will be released soon.

The model costs $0.07 per million input tokens and $1.68 per million output tokens.

With the new release, DeepSeek brings many features seen with other AI model providers, including better inference capabilities, Chandrasekaran said.

"The hybrid inference architecture is quite noteworthy here, offering faster processing and enhanced agentic capabilities, but none of this addresses the safety and regulatory risks associated with DeepSeek," he said.

Shimmin said the release also reinforces the idea that instead of taking a giant stride forward, AI model makers are investing in making existing models more effective and efficient.

"[The] subtle investments are critical for making these models enterprise-grade out of the box," he said.

He added that the small changes in making each model efficient are essential for enterprises deciding between different models and what works for their tasks.

"We're seeing a maturation of the AI space that we as practitioners in the enterprise would ask for and want," Shimmin  said. "We're getting a refinement of capability and improving enterprise-grade concerns."

Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.