Google Cloud announces 'Gemini agent' as universal agent for work

Google Cloud has announced Gemini agent, a cloud-based AI agent designed to act as a universal worker across enterprise tools rather than waiting for a sequence of detailed instructions.

The announcement came at Gemini at Work 2026 on 8 October 2026. Google describes Gemini agent as a single agent for work that can take objectives, decide how to approach them and carry out knowledge-work tasks across the Gemini Enterprise app, Google Workspace and third-party services.

What Gemini agent is meant to do

Google says the agent can answer questions, create media, write and run code, and handle multi-step jobs. Its proposed workflow is deliberately outcome-led: a request such as turning a project update into a slide deck can be handed over as a delegatable task instead of being broken into every individual step.

The same approach is intended to cover coordination work. Google gives the example of asking Gemini to arrange a meeting with a team of regional event leads. The agent can infer the relevant people from Workspace context, check calendars and begin coordinating with internal or external participants.

Gemini agent can also operate as a team member, Google says, working for a group or a defined organisational role such as a finance analyst. That positions it as more than a chatbot window: the agent is supposed to keep working towards a broader objective while managing the smaller jobs needed to reach it.

Models, sub-agents and memory

Under the hood, each job is routed to the model Google considers the best fit. Google says that currently includes Gemini and Claude model families, with other private and open models planned for the future. The stated aim is to balance quality with cost rather than send every request to one fixed model.

Gemini agent can also use sub-agents for multi-step work, coordinating parallel and sequential tasks that may run for hours or days. Google says the cloud-based system uses one shared set of memories, context and a personalisation graph. Those memories cover the current session, knowledge it builds, the procedures it learns and a record of previous work.

Where it will work

Within Gmail, Docs, Sheets, Slides and Chat, users will be able to invoke Gemini with @Gemini. Google also lists command-line access, Slack and Microsoft 365 among the channels it wants the agent to reach, while an API will let developers integrate it as a headless agent without a dedicated user interface.

The Gemini agent is currently in private preview. Google says wide availability is coming soon for Workspace customers on selected Business and Enterprise plans; the announcement does not provide a firm public launch date.

Our opinion

The interesting part of Gemini agent is not the promise of another assistant that can draft text. It is the attempt to turn enterprise context into an operating layer: the agent is meant to remember how work is done, choose among models and tools, and keep a task moving after the initial request.

That ambition also makes the preview important to watch. An agent that can infer colleagues, access calendars, run code and act across services will be judged less by its demos than by its permissions, audit trail and ability to explain what it did. Google has supplied the broad architecture, but the private preview will have to show whether the convenience of objective-based work outweighs the governance burden.