On 8 October 2026, at its Gemini at Work event, Google Cloud introduced the Gemini agent, which it calls a single, universal agent for work. You give it a goal from one prompt box, and it answers questions, does knowledge work, creates images and media, and writes and runs code to get there. The most striking part isn't any single feature. It's the idea that a company can give an AI agent its own identity, with an email address, a calendar and a place in the company directory, and treat it like a junior colleague.
What the Gemini agent does
Google's pitch, in the words of Google Cloud chief executive Thomas Kurian, is that "you give it objectives, not instructions." The agent is designed for long jobs: Google says work that takes hours or days keeps running after you close your laptop. It can break a big task into smaller ones, start temporary sub-agents to handle them and schedule work for later.
It sits inside the tools most Google customers already use, including Gmail, Drive, Docs, Sheets, Slides, Chat and Calendar, and it is also reachable from the web, iOS and Android, Microsoft 365 and Slack. Through a tools registry it connects to business systems such as Salesforce, Jira, Confluence, ServiceNow, Microsoft Teams, Git, BigQuery, Snowflake, Databricks and Postgres, plus any MCP server, the open standard many AI tools now use to plug into other software.
AI coworkers with their own email address
The feature people will talk about most is what Google calls coworker agents. These are agents with their own identities, including an email address on a dedicated agents subdomain of your company, a Workspace account, a calendar, Drive storage and an entry in the directory. In practice, you could have an agent that people email with requests, invite to meetings or share files with, just like a person.
That sounds risky, and Google spends a lot of its announcement on control. Each agent gets least-privilege permissions, meaning only the access it needs, and those permissions are approved by the organisation's security administrators. Every action is written to an audit trail. Companies can set real-time spending caps, and if a cap is hit, that project's agent pauses. Google also says the agent can orchestrate across its own Gemini models and Anthropic's Claude models, which is a notable admission that no single model family is best at everything.
Who can use it, and what is still unclear
This is a product for companies, not individuals. Google's announcement doesn't give availability dates or pricing, and launch coverage describes the agent as being in private preview for a select group of customers. Versions tuned for financial services and legal work are in preview, with government, healthcare and retail coming later. If you are a Workspace customer and haven't heard from Google, you probably don't have it yet. Treat any price you see quoted online as unconfirmed until Google publishes one.
The Gemini agent arrives in the same month as Anthropic's Claude Duet and OpenAI's push into always-on agents. The approaches differ. Duet is a personal assistant for individual professionals that checks in with you before acting. Google's agent is built for organisations, managed by IT, with identities and audit trails. Both point the same way: the next phase of workplace AI is agents that carry out work, not chatbots that answer questions.
Other Google changes worth knowing about
If you use Gemini personally, the change that affects you soonest is that Gems become skills from 17 November 2026. We cover what carries over and what breaks in our guide to moving custom GPTs and Gems to skills. Workspace business and enterprise accounts follow in March 2027.
How to prepare your team
You don't need access to get ready. An agent is only as good as the instructions and access it is given, so start by writing down two or three processes your team repeats every week, step by step, with the systems each step touches. Those notes become the objective you hand to an agent later, whichever company's agent you end up using.
Next, find out who would approve an AI agent's access at your company, because in Google's model, security administrators decide what each agent can touch. Having that conversation before the product arrives saves months. And keep your instructions portable. Writing your workflows as skills in the open Agent Skills format means they can move between Gemini, Claude and ChatGPT. Our Stacks show which tools teams combine for common workflows, and the Gemini page in Learn shows what it does best today.