Google has presented Gemini Agents, a new generation of artificial intelligence agents designed to work autonomously from a goal, rather than from detailed step-by-step instructions. The announcement, reported by TechRadar Pro, describes agents that run on Google Cloud infrastructure and can therefore stay active even when their user is no longer in front of their screen. For professional use, they will be able to have their own email address, their own calendar and Drive storage space.

Before imagining what this changes, some sorting is needed: part of the announcement describes a product in preview, while another part is clearly a matter of corporate vision.

What Google describes

The principle is one of delegation. Instead of asking an AI to write a paragraph or summarise a document, the user would entrust it with a complete assignment – for example preparing a file, producing resources or organising a meeting – and the agent would work out the steps itself. Google speaks of a central agent capable of answering questions, creating documents, generating resources and organising meetings, among other tasks.

Technically, the key point is execution in the cloud. The agent does not depend on the user’s computer or phone: it runs on Google’s servers, which allows it to carry on with long tasks in the background. It is this architecture that makes the idea of an “always-on” agent possible.

Google also stresses openness. According to the announcement, Gemini Agents will be compatible with Anthropic’s models, and other models should be supported in the future. The service is to work with third-party platforms such as Microsoft 365 and Slack, as well as on devices running iOS, Android, macOS and Windows. Thomas Kurian, CEO of Google Cloud, sums up the ambition as that of a “durable platform that works everywhere, executes any task, supports model choice and applies the controls enterprises need”.

Google is not alone in this field: TechRadar Pro points out that the announcement comes shortly after similar presentations by Meta, with Muse, and OpenAI, with Dots.

Available, announced or envisaged?

This is the most important distinction for the reader. What is established: Gemini Agents is not yet generally available. Only certain customers have access to a private preview. No general availability date is mentioned in the reported announcement, and more details on uses and features are expected as this launch approaches.

What is announced but not yet verifiable: compatibility with models other than Anthropic’s, extended integration with third-party tools, and each agent’s own email address and calendar. These are promised features, whose actual behaviour can only be assessed once the service is opened more widely.

What belongs to the outlook: the idea that agents will become genuine “digital employees” capable of carrying out long assignments in a largely autonomous way. This is the vision Google is putting forward; nothing in the announcement yet makes it possible to measure the reliability of these agents on complex tasks.

The practical questions for users in Switzerland

The announcement says nothing about Switzerland. The first question is therefore simple: when and under what conditions will the service be accessible to Swiss businesses and individuals? Is the private preview open to customers here, and will the features be the same as elsewhere?

Next comes the question of data. An agent that has its own email inbox, its own calendar and a Drive space, and that can connect to Microsoft 365 or Slack, by design has access to a large amount of internal information. Where is this data processed and stored? What contractual guarantees apply when the agent calls on a third-party model, such as Anthropic’s? These points are not detailed in the announcement.

Finally, there is the question of control. An agent that acts in the background, without its user being logged in, and that can send emails in its own name, raises very practical questions: which actions require human approval? How can one trace what the agent has done, and correct an error? Who is accountable for a message sent or a meeting organised by an agent? Thomas Kurian refers to the “controls enterprises need”, without their content being specified at this stage.

What this changes, for now

In the short term, little for most users, since the service remains in private preview. Above all, the announcement marks a direction: the major AI players, Google as well as Meta and OpenAI, are shifting their offering from the assistant that answers to the agent that acts. For organisations considering this type of tool, the moment is probably better suited to preparing the questions – access governance, traceability, data location – than to adoption.

Sources