Google Cloud announced its Gemini agent on Oct. 8, describing it as a cloud-based work assistant that can use company tools and keep assigned jobs running for hours or days after a user closes a laptop. The company also says it can choose between Google's Gemini and Anthropic's Claude models for different jobs.
An AI agent can plan steps and use connected software to pursue an assigned outcome, rather than only answer one prompt. Google describes the same Gemini agent handling chat, knowledge work, media creation and code from one interface and API. For larger jobs, it can create temporary subagents and coordinate their work in sequence or in parallel.
Google says the agent runs in the cloud and maintains one set of memory and context across the devices and channels through which it is accessed. That is the company's explanation for how work can continue after a laptop closes. Google names On as an early tester of its model-selection feature, but the post gives no general-availability date for the full agent.
The agent is designed to carry its context across company tools. Google lists Workspace apps such as Gmail and Drive alongside Microsoft 365, Slack, Salesforce, ServiceNow, databases and custom connectors. That could let a task move from finding material to preparing a document or running code without a person copying context between separate assistants.
Google says the agent can route jobs across its Gemini models and Anthropic's Claude models today, with other open and private models planned later. The agent's saved context, tools and company rules are meant to stay in place when the underlying model changes.
Google also describes an individual identity for each agent, administrator-approved permissions, action logs and a sandbox with a network gateway that enforces company policies. These controls matter when software can keep working after a user signs off, access shared files and run code. The announcement describes Google's design, not an independent security audit.
Google's launch post offers no third-party evaluation of the agent's long-running, multi-app performance; its customer examples and control descriptions do not show how reliably it would work in other organizations.
Sources
- Google Cloud: Gemini at Work 2026 and the Gemini agent
- Google's event announcement
- Illustrative photo: Front of server racks at NERSC, photographed by Derrick Coetzee, released under CC0 1.0. The racks are not Google's Gemini systems.
