Google Cloud launches a Gemini agent for work across business apps

Why it matters

Google is selling an agent layer that can work across workplace suites, enterprise systems and competing AI models. Its value will depend on whether persistent context and permissions let it complete real cross-system tasks reliably.

Google Cloud CEO Thomas Kurian (@ThomasOrTK) announced a Gemini agent for enterprise work on October 8th, pitching it as one system for answering questions, creating content, writing code and carrying out longer assignments across business software. The announcement came at Google Cloud's Gemini at Work event and extends Google's presentation beyond a chat interface: the agent can keep running in the cloud after a user leaves, coordinate sub-agents and return work inside the applications employees already use.

https://x.com/ThomasOrTK/status/2108245829530325307

Kurian's pitch reflects his long focus on enterprise software. He joined Google in 2018 after 22 years at Oracle, where he was president of product development, and previously worked at McKinsey, according to Google Cloud's leadership biography. His remit at Google Cloud has been selling technology into organizations that already rely on sprawling software stacks. The new agent is designed to work across those stacks rather than require every user to start in a Google app.

Google says Gemini can connect to tools including Microsoft Office, Teams, Slack, Confluence, Git, Jira, Salesforce and ServiceNow, along with databases such as BigQuery, Databricks, Postgres and Snowflake. It can also run from a command line or inside third-party applications, and can operate without a dedicated interface. These integrations let the agent take on tasks that cross departments and systems, which an agent limited to a company's documents or a single software suite may not be able to handle.

Google says Gemini maintains memory and context across devices and channels, and can continue work lasting hours or days after a user closes a laptop. For multi-step assignments, it can create temporary sub-agents with distinct identities and coordinate tasks in parallel or sequence. Teams can also set up persistent "coworker agents" with defined roles, storage and company email identities. Google says those agents receive only the context shared with them.

Google is positioning Gemini as an interface between employees and workplace software, with access to systems from Microsoft and other vendors. The company also says Gemini can select among its own models and Anthropic's Claude models for different jobs, with private and open models to follow. Customers could keep workflows and organizational context in one place while switching the model underneath them. Google says this can help control costs, but the announcement supplies no independently tested comparison of quality or savings.

The product arrives as Google argues that enterprise AI has moved beyond pilots. In its event post, Google said nearly 80% of Google Cloud customers use its AI products and nearly 90% of Fortune 100 companies use Gemini Enterprise. Those figures describe Google's existing AI business, not adoption of the newly announced agent. Google also said that nearly 500 customers each processed more than one trillion tokens during the preceding year. Google did not report usage of the new agent.

Kurian's event-day X thread summarized the product around a single prompt box and a shared layer of business context. Google's longer announcement adds inline work in Gmail, Drive, Docs, Sheets, Slides, Chat and Calendar, plus security controls such as agent identity, role-based permissions and activity logging. Those controls will determine whether companies let software act across sensitive systems rather than simply draft a response for an employee to review.

Google is betting that enterprises will accept a general-purpose agent that moves across competing software and models if it can remember how their organization works and complete tasks under company permissions. The test is whether that context and governance hold up across real workflows, not how many capabilities fit behind one prompt box.