From Las Vegas, at Google Cloud Next, the American giant is announcing a host of new AI developments to become the agentic reflex on a corporate scale.
“The question is no longer ‘can we build an agent?’ but how to manage thousands of them? At Google Cloud Next in Las Vegas, where the JDN is located, Alphabet boss Sundar Pichai assumes a strong thesis: tomorrow, the company will no longer be governed solely by its employees, but by thousands of AI agents orchestrated continuously. In this future, Google wants to be the default reflex. No longer the gateway to the web, but that of agents. From the agentic platform to the proactive digital workplace, including hardware, data and security, the cloud giant is aligning its entire stack to become the reflex for businesses in the AI era.
Gemini Enterprise Agent Platform: agentic at scale
Google is betting everything (or almost) on Gemini Enterprise Agent Platform, its new autonomous platform for building, deploying, orchestrating and governing complete fleets of AI agents in production. A new product focused around four pillars: build, scale, govern, optimize. The message from Thomas Kurian, CEO of Google Cloud, is clear: companies no longer want to cobble together isolated agentic POCs, they want a unified platform to manage thousands of agents with the same rigor as their critical systems.
Gemini Enterprise Agent Platform arrives with an arsenal of bricks designed to industrialize the construction of agents:
- Agent Studio : a low-code interface to design, test and publish agents in natural language
- Agent Developer Kit (ADK) : framework (for developers) to orchestrate the logic between agents
- Agent Registry : a unique directory referencing each agent and each internal tool (a very important tool for the mapping of AI systems required in particular by the AI Act in Europe)
- Marketplace Agent : a marketplace to deploy third-party agents in one click
On the protocol side, Google is pushing MCP everywhere. All Google Cloud services (BigQuery, Kubernetes, AlloyDB…) now have their MCP servers. Agents can thus interact natively with the entire Google stack.
But it is on governance that Google hits hardest, and this is probably where the real differentiation against Bedrock or Azure comes into play:
- Identity Agent gives each agent a unique cryptographic identifier, distinct from that of the human user, with traceable and auditable authorization policies.
- Gateway Agent acts as a control tower for agents by monitoring all operations (in particular, it makes it possible to detect and act on attacks by prompt injection, tool poisoning and data leaks)
- Anomaly Detection Agent makes it possible to detect “reasoning drift”, the subtle deviations of an agent from its initial task
- With Secure sandboxes publishers can use virtualized environments for agents to execute code or computer use in a secure environment
- Memory Bank offers persistent memory management over several months (on the principle of OpenClaw)
- Agent Observability allows a human to follow in real time what agents are doing in production
- Agent Simulation allows you to test an agent before it is put into production by confronting it with thousands of simulated scenarios in stress test mode
For Google without this advanced governance layer, agentic at scale is untenable for a large enterprise.
Google Workspace redesigned around agentics
Google pushes agents in all its products and the workspace suite is no exception. The main new feature: Workspace Intelligencea semantic layer that connects Gmail, Drive, Docs, Sheets, Meet and Chat in a single context that agents can use. Concretely, all of a user’s Workspace data (emails, files, chat threads, meeting notes, calendar events) are indexed together and searchable by agents as a single knowledge base. Google is also adding Ask Gemini to Google Chat to query your entire Workspace without switching tabs, and Google Drive Projectsa space that automatically organizes files and emails around the same project (common context). Finally, Docs, Sheets and Slides are accessible from a mode Canvas allowing you to create and edit documents directly from Gemini Enterprise.
Google is seriously strengthening Gmail with two bricks already deployed to the general public but which are coming to the business suite. AI Overviews in Gmail will automatically summarize long email threads. Even more interesting, AI Inbox will reorganize the inbox proactively. Gemini will proactively sort emails by real priority (not just by date), group related topics and flag what is awaiting a response. The idea is to move from a chronological inbox to an intentional inbox.
A new vision for TPU
To support this transition to agentic AI, Google is pushing its efforts into hardware. The publisher announces its eighth generation of TPU, with a new strategy: two separate chips for two uses. The 8t TPU is dedicated to training models. It is possible to connect up to 9,600 chips in a single superpod. The TPU 8i is designed for inference. It is possible to connect 1,152 chips per pod and Google promises a performance/price ratio 80% (yes, 80%) higher than the previous generation.
Google is thus approaching the AWS model, which has dissociated its Trainium (training) and Inferentia (inference) chips for several generations, while Nvidia has kept a unified architecture with Blackwell. Enough to interest companies who want to train their own models by truly optimizing costs, whether for training or inference.
Agentic, traction of tomorrow for Google?
Beyond the agentic layer, the entire Google Cloud stack is reinventing itself: data architecture, cybersecurity, storage, networking. With this flood of announcements (which it is impossible for us to deal with in a single article), Google aims to become the default reflex for the agentic company. The traction on generative AI is real: 16 billion tokens (via API) are processed per minute currently, compared to 10 billion the previous quarter.
But moving from content generation to the orchestration of autonomous agents is a leap of a completely different nature. And in this area, Google is not going alone: Microsoft deploys its agents in a 365 suite already well established on the market, AWS is moving forward with its own agentic ecosystem, and more agile players like Anthropic or OpenAI are building their own orchestration platforms. The real difference? Google may be the only one to simultaneously align the three pillars that agentic truly demands: data, mature cloud infrastructure, and cutting-edge generative models.
