Google wants the next member of your team to be an AI agent. At its Gemini at Work event on October 8, Google Cloud unveiled Gemini agent, which the company describes as a "universal agent for work". The idea is easy to explain and hard to execute: a single prompt box that handles questions, knowledge work, content creation and coding, instead of a collection of separate tools for each task.
Inside the tools you already use
According to Google, the agent is available in the Gemini Enterprise app and inside Google Workspace, meaning Gmail, Docs, Sheets and Calendar. The more surprising detail is that it can also be used from within Microsoft 365 and Slack. In doing so, Google openly acknowledges the reality most organizations live in: they do not use a single vendor's tools, and an agent that only works inside one system will stay limited.
"Coworkers" with their own permissions
The most interesting feature is the ability to create "coworker agents". These agents act as team members, with their own email address and access only to the information they are explicitly given. In practice, you can set up an agent that handles customer requests, prepares weekly reports or tracks project tasks, and communicate with it like any other employee.
According to an analysis by research firm Futurum, the launch also includes a broad control layer: a separate identity for each agent, permissions passed through OAuth, a log of every action the agent takes, an isolated runtime environment and real-time spending limits. These are exactly the questions that worry IT and security teams when it comes to agents acting on an organization's behalf.
Not just Gemini
Another notable detail: the agent chooses the right model for each task, and it currently runs both Google's own models and Anthropic's Claude models. Google said more models will be added later. The decision to include a competitor's model shows how far the enterprise market has become about choice: customers want the best model for each task, not necessarily the model of the vendor they buy from.
Alongside the general version, Google is previewing versions tailored to financial services and legal work, and promises versions for government, healthcare and retail later on.
What to check before adopting
For organizations evaluating the tool, the key question is less "what can the agent do" and more "what is it allowed to do". An agent with an email address and access to documents can save hours of work, but it can also email the wrong person or expose an internal document. Start with well-defined, low-risk tasks, give the agent only the minimum permissions it needs, and review its action logs regularly. The technology is moving fast, but responsibility for the outcome still rests with the people running it.
