These four new features of Google Cloud Next 2026 are perhaps those that will have the most impact on your AI projects.
If Gemini Enterprise Agent Platform is on everyone’s lips in Las Vegas during the 2026 edition of Google Cloud Next, several announcements went more unnoticed… but caught the attention of the JDN, present on site. Cloud, cybersecurity, data… Agentic AI now offers clear transformative power for the majority of organizations. Focus on the 4 new features which, in our opinion, will have a real daily impact on the companies that adopt them.
1. Multi-cloud AI
To be truly relevant, generative AI requires access to data. However, company data is, even more so in 2026, often distributed among different cloud providers (this is especially true for large organizations). And it is very complex to set up agentic applications using data stored in several cloud providers. To face this reality, Google presents its Cross-Cloud Lakehousea data infrastructure designed natively for agents.
The principle is simple: rather than forcing companies to bring everything to the same place, Google Cloud now allows its AI agents to fetch data where it is (whether on AWS, Azure or Google Cloud) without having to move it. On stage, Karthik Narain, CPO of Google Cloud, summed up the promise: “No more data transfer, no more dependence on providers, room for freedom.”
2. Deep research agents for business data
Second discreet but equally important announcement: deep research agents. Google is launching two new versions of its autonomous search agent, deep research and deep research Max, now accessible via the Gemini API. With a simple API call, an agent can chain a web search, query BigQuery and synthesize internal documents to produce a comprehensive, sourced analysis. The agent no longer works on raw data, it reasons on an enriched business context, with mapped entities, relationships between sources, and verifiable citations. Presented as a native component of Gemini Enterprise, the tool inherits all the governance and access policies already in place in the company.
3. An agentic red team
With AI, “the time between an initial intrusion and the handing over to a second group of attackers has fallen from eight hours to 22 seconds in three years,” warns Francis deSouza, COO of Google Cloud. The cause? AI. It now allows malicious groups to automate and accelerate their attack chains at unprecedented speed.
To face this new reality, Google is banking on… AI. Three specialized agents are presented: a “Red Agent” which continuously tests the IS offensively, a “Blue Agent” dedicated to defensive surveillance and a “Green Agent”. The latter applies end-to-end remediation and automatically goes back to the responsible line of code, generates a fix and sends it directly as a pull request to the developer concerned, without human intervention. The three agents form an autonomous defense loop.
4. Multimodal RAG becomes reality
Latest announcement, on the sidelines of Next: the arrival in general availability of Gemini Embedding 2. The model was unveiled a few weeks ago, but now becomes accessible to everyone via the Gemini API and in Vertex AI. While foundation models like Gemini 3.1 get all the attention, embedding models are just as important. They are the ones that make it possible to vectorize the data to make them searchable by an LLM. However, until now, the vast majority of embedding models on the market only worked on one modality at a time.
With Gemini Embedding 2, Google allows you to vectorize any type of content (audio, video, text, etc.) within the same model. Concretely, a RAG can now become natively multimodal. And because all modalities are projected into a common vector space, the consistency and precision of the AI’s responses are greatly improved.
