Google Cloud’s Gemini agent is one workplace agent, still in private preview

At Gemini at Work on October 8, 2026, Google Cloud CEO Thomas Kurian introduced the Gemini agent as one workplace agent and one API. In a post adapted from his keynote, he calls it a single universal agent for questions, knowledge work, images and media, and code, all from one prompt box. The Verge, reporting the same day, says that agent is available only to enterprise customers in private preview, inside the Gemini Enterprise app. Neither account states a price or a general-availability date.

The operating idea is delegation, not a longer chat. Kurian says users give the agent objectives, not instructions: it plans the work, uses skills and tools, connects to company systems, and returns finished work in the documents, inbox, and developer environments people already use. They can assign work, schedule tasks, or have it respond to events. It can act as a personal assistant or on behalf of a team or a role, such as a project manager or a finance analyst.

Because it runs in the cloud, Google says one set of memories, context, and a personalization graph follows the user across devices and channels. There is no re-brief when someone switches machines. Work that takes hours or days keeps running after the laptop is closed and is still there when they come back.

Where the same agent shows up

Google lists access from the web, iOS, Android, Windows, and Mac, plus the command line, Google Workspace, Microsoft 365, and Slack. It can also be integrated into third-party applications and run headless, without a dedicated interface. The Verge’s account matches that spread: Workspace apps, mobile, desktop, the web, Slack, and Microsoft 365, with the same cloud context across connected devices.

Inside Workspace, Kurian says the agent carries the same memory, skills, and controls into Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, working in the email thread, the document, or the chat space. Google describes three modes there: personal assistance that already knows calendar, team, and projects; proactive suggestions to delegate a task, such as turning a manager’s request for a project update into a slide deck; and a coworker agent for the whole team.

That coworker is a persistent role, not a one-off chat. Google says coworker agents get their own @agents.company.com email, their own storage, and only the context a person or team shares. In Workspace, the agent can receive its own account, including email, calendar, Drive, and a company-directory presence. Colleagues add it to a Chat space or @mention it. It acts under its own identity, and access follows the sharing and membership the team already uses.

The agent can also spin up temporary, job-specific sub-agents, each with its own identity, and coordinate parallel or sequential steps that run for hours or days.

The agent routes the model

Kurian’s sharper product claim is that Gemini is the agent and the model underneath is a separate choice. He says each job runs on the model that fits it, orchestrating across Google’s Gemini family and Anthropic’s Claude today, with other private and open models planned. The argument is cost as well as quality: the best model for a task is not always the largest, and keeping context, skills, and data on the agent means they stay put when the leading model changes. The Verge is narrower here. It says Gemini automatically uses the model that best fits the job, and does not name Claude.

Google also describes Smart Routing, which triages workloads onto the model it says delivers the needed performance at the lowest cost, and real-time spend caps in the Cloud Billing Console. If a project cap is hit, that project’s agent pauses until someone resumes it. Tracking is per project, so costs can be charged back to a department.

Named connections include collaboration tools such as Confluence, Microsoft Office, Teams, Slack, and Workspace; development tools such as Git and Jira; Salesforce and ServiceNow; BigQuery, Databricks, Postgres, and Snowflake; desktop files; and any Model Context Protocol server inside or outside the company network. Teams can publish tools and skills to company registries. Skills are reusable sets of instructions, knowledge, or workflows. Google says the agent keeps four kinds of memory: session memory for the task in front of it, semantic memory built from documents and other agents, procedural memory for how a job gets done, and episodic memory of what it has done before.

Industry packs are a separate preview, not a general release of those verticals. Kurian says specialization is in preview for financial services and legal, with skills, tools, connectors, and domain knowledge for each, and coming soon for government, healthcare, and retail. The financial-services version is described as drawing on FactSet, LSEG, S&P Global, SEC filings, and a company’s own repositories, with confidence scores, methodologies, data lineage, and source citations. The legal version is described as inheriting matter-level permissions and ethical walls from document systems such as NetDocuments and iManage.

Governance is part of the same pitch. Google says every agent gets its own identity, least-privilege permissions approved by security administrators, and an audit trail attributed to the agent rather than a person. Agents execute inside an Agent Sandbox, and traffic in, out, and between agents passes through an Agent Gateway that enforces policy in real time.

For buyers, the useful line is preview versus rollout. Read Kurian’s Gemini at Work post for the product scope, then The Verge’s report for the access limit. Enterprise customers who want to try it should ask their Google Cloud account team about private preview, and should not plan a migration around a ship date or a price that these accounts do not give.

Where the same agent shows up

The agent routes the model

Sources