Gemini Agent: Google Cloud Turns Gemini Into One Universal Agent for Work (With Its Own Email and Calendar)
On October 8, 2026, at Gemini at Work 2026, Google Cloud unveiled the Gemini agent: a single universal agent for work that lives inside Gmail, Docs, Slack and Microsoft 365, takes objectives instead of instructions, and even gets its own inbox and calendar.
What Was Announced on October 8, 2026
Google Cloud used its annual conference to present a product, not an upgrade: an agent that answers questions, does knowledge work, creates media, and writes and runs code from a single interface and a single API.
Gemini at Work 2026: One Agent, One API
The message is consolidation: one surface with the same memory, the same skills and the same controls everywhere. For people who build software, the relevant part is not the chat window but the fact that work comes in through a single API.
Where the Agent Lives: Workspace, Microsoft 365, Slack and the Command Line
It does not ask you to change tools: it works in the Gemini Enterprise app, in Google Workspace (Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar), in Slack and Microsoft 365, and on the command line. The promise is that the channel does not matter: it keeps context and returns finished work wherever you asked.
Read also
Private Preview for Enterprises: What That Actually Means
It is in private preview for selected Gemini Enterprise customers, with industry versions in the same phase. These are announced capabilities, not a product anyone can buy today: if your company is not a customer, there is no purchase button.
From Conversational Assistant to Agent With an Identity
The underlying shift is that Google stops selling an assistant that answers and starts selling an agent that executes. And an agent that works needs identity, permissions and traceability, so the decision lands on IT.
What the Gemini Agent Is and How It Differs From the Assistant
Tasks That Keep Running After You Close the Laptop
An assistant lives inside the session: close the window and it is over. The agent is persistent: it takes on tasks that run for hours or days in the background and returns finished work, not a draft waiting for review.
Its Own Identity: the Agent's Email, Calendar and Storage
As announced, coworker agents get their own Workspace account, with a dedicated email address, calendar and Drive storage, and they only reach the context the team gives them. It stops being a feature inside an app and becomes another actor in the organisation.
Temporary Sub-Agents: How Multi-Step Work Gets Organised
For long tasks it can spin up temporary sub-agents, each with a specific responsibility, while it coordinates the steps in parallel or in sequence. That is multi-agent orchestration without the jargon: split the work, then put it back together.
You Give It Objectives, Not Instructions
The agent plans, uses the available tools, connects to company systems and brings back something finished. It is a change in the contract: less micro-management, more verifying the outcome.
The Numbers Google Put on the Table
Google said that over the past year nearly 500 Google Cloud customers each processed more than a trillion tokens, that close to 80% of its customers already use its AI products, and that about 90% of the Fortune 100 uses Gemini Enterprise. These are company claims, not audited data: they measure volume and adoption, not quality.
What We Still Don't Know
No Price, No General Availability Date, No Waitlist
No price was published, no general availability date, no waitlist. Any figure you see is made up. The API documentation does clarify that agent billing runs on token consumption and tool usage, not per seat.
Model Routing and Why It Is a Platform Decision
Coverage confirms the agent can spread work across Google models and Anthropic models. It matters for three reasons: cost per task changes with the model, evaluation is no longer about a single model, and you need to know where data ends up at each step.
Permissions, Audit Trails and the Lifecycle of an Agent
It is still open how its actions get audited, what happens when an agent with its own inbox is retired, or how inherited permissions behave. Those are the questions any systems owner will answer before granting real access to their data.
What It Means for Developers and Technology Decision-Makers
If the Agent Does the Work, Your Data and Actions Are What Matter
Your interface stops being your edge: what counts is that your data can be read and your actions run safely. The realistic pattern for a small team is to be the source or the action the agent needs, not to compete with its window.
MCP and Tools as the Real Integration Layer
That is where the Model Context Protocol (MCP) comes in: a standard for exposing your tools and data to a model without writing a separate integration per vendor.
What Gets Commoditised: Forms, Templates and Summaries
Anything that is a form, a template or a summary falls inside the reach of a general agent. The healthy response is moving toward what an agent cannot do alone: your own data, hard integrations, legal responsibility and specific domain knowledge.
New Decisions: Identity, Least Privilege, Cost per Task and Rollback
Four decisions that did not exist before: what identity the agent has, what least privilege it runs with, what each task costs and how to roll back something it already did.
The Week's Map: Three Announcements, Three Different Grounds
In recent days Anthropic moved its budget model, OpenAI brought GPT-6 with its Intelligent UI to the consumer chat, and Google made its move in the enterprise. They are not the same thing: that one is how an answer is presented; this one is an agent with an identity that executes work in your systems. On models, the reference is still Gemini 4 Argon.
What Comes Next
When price and general availability arrive, how Microsoft answers with Copilot, and whether the agent with its own identity becomes the standard. If it does, the advantage will not be in having an agent but in having permissions, auditing and lifecycle sorted out.
Conclusion
What a company gains is finished work instead of drafts; what it bets is one more layer of proprietary platform and a bill tied to consumption. If you build software, the useful question is not whether an agent will replace you, but whether your data and your actions are ready for someone to use them. The blog will keep following this layer in the coming weeks.
