Google Cloud launches Gemini agent, a multi-model enterprise agent that runs Gemini and Claude, in private preview

Google Cloud announced the Gemini agent at its Gemini at Work 2026 event on October 8. The company describes it as a "new single, universal agent for work" for enterprise customers. It currently runs on both Google's Gemini models and Anthropic's Claude models. It is in private preview.

What Google announced

The Gemini agent is available in the Gemini Enterprise app, Google Workspace, and third-party services. Google says users can "give it objectives, not instructions." Its stated capabilities are answering questions, handling knowledge work, creating media, and writing and running code.

Google gave two examples:

• Delegation from email: If a manager emails asking for a project update as a slide deck, Workspace Intelligence flags it as a delegatable task and offers a single-click handoff to Gemini.
• Meeting coordination: Asked to schedule a meeting with the usual regional event leads, Gemini infers the attendees from chat space membership and the previous event thread. It checks calendars and starts an email thread to find a time, including with external participants. No names or email addresses are needed.

Google also positions the agent as a "team member" that works on behalf of a group, such as a project manager, or for a specific role, such as a finance analyst.

Architecture

Model routing. According to Google, the agent "runs each job on the model that fits best." Today that covers the Gemini and Claude model families. Google says support for other private and open models will follow. It claims this approach delivers "optimal quality and lower your costs." Google did not name specific model versions or publish cost comparisons.

Sub-agents. The agent uses sub-agents for multi-step tasks and connects to third-party tools. It coordinates parallel and sequential workflows that Google says can run for hours or days.

Memory. The agent runs in the cloud with "a single set of memories, context, and one personalization graph." Google lists four memory types:

• Session: context for the current task, even across multi-day runs
• Semantic: a structured knowledge base built from documents, conversations, and other agents
• Procedural: how jobs get done, including skills the agent writes for itself
• Episodic: a record of everything it has done before

Access

Users can invoke the agent by mentioning @Gemini in Gmail, Docs, Sheets, Slides, and Chat. Google lists the command line, Workspace, Slack, and Microsoft 365 as examples of supported channels. An API lets third parties embed it as a "headless agent," meaning it needs no dedicated interface.

Availability and pricing

The Gemini agent is in private preview. Google says it will see wide availability "soon" for Workspace customers on select Business and Enterprise plans. Pricing, context window, benchmark results, and a general availability date have not been disclosed. The specific plans covered were not named in the source.

What this means

The notable point is the model-agnostic design. Google is putting a competitor's models, Anthropic's Claude, inside its flagship enterprise agent and treating the model as a swappable component. The product is the orchestration layer: memory, routing, sub-agents, and integration with Workspace data.

The cross-platform reach matters too. Support for Slack and Microsoft 365 suggests Google wants the agent to operate where enterprise work already happens, not only inside Workspace. The headless API supports this, because developers can embed the agent in their own products.

The cost-reduction claim is unverified. Routing each job to the best-fit model could lower costs, but Google has published no pricing or benchmarks. Until the preview opens up, the persistent memory, self-written procedural skills, and multi-day autonomous runs remain vendor claims. Enterprise buyers should test them against their own data and governance requirements.