Google launches Gemini agent for work in private preview, introducing a cloud-based assistant that can plan assignments, use enterprise tools and deliver finished work across office and developer applications. Google Cloud announced the product on October 8 as the central release of its Gemini at Work event.

 

The agent spans Google Workspace, Microsoft 365 and Slack while retaining task context in the cloud. It can also route work between Google Gemini and Anthropic Claude models, a notable departure from enterprise assistants that keep every step inside one model family.

 

The initial release combines four capabilities:

  • Persistent context across apps and devices
  • Model routing between Gemini and Claude
  • Long-running work through delegated subagents
  • Company-managed coworker agents with separate identities

 

Google Launches Gemini Agent for Work in Private Preview

Google describes Gemini agent as one interface for questions, task execution, content creation and software development. It is available through the web, mobile apps, Windows, macOS and a command-line interface, with integrations that place it inside the software employees already use.

 

The product is not a general public release. The Verge reported that it is currently in private preview for enterprise customers, so Google has not yet provided broad availability dates or standard pricing. That boundary matters because the most ambitious functions depend on customer data, permissions and administrative controls.

 

Unlike a chatbot that waits inside one window, the agent can continue working after the user moves to another device. Google says its cloud memory preserves relevant context, preferences and prior decisions, allowing a task started in email to continue in a document, spreadsheet or development environment.

 

The agent can also run without an open user interface. For larger assignments, it may create subagents that work for hours or days, coordinate their contributions and return completed material. That design targets research, analysis and multi-step production work rather than isolated prompts.

 

How Gemini Agent Routes Tasks Across Workspace and Claude

Google is positioning orchestration as a core feature. The agent can select a model according to the task, using Google’s Gemini family and Anthropic’s Claude models at launch. Google says support for additional models is planned, but it has not named them or committed to a release schedule.

 

The tool registry is similarly broad. Google lists connections to Gmail, Docs, Sheets, Calendar, Confluence, Microsoft Office, Teams, Slack, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres and Snowflake, along with desktop files and Model Context Protocol servers.

 

That breadth gives the agent a potential route from understanding a request to completing it. A finance employee could ask for analysis that draws from a warehouse, checks supporting documents and returns a spreadsheet. A developer could assign a coding task that uses repositories, issue trackers and a command-line environment.

 

The approach also creates a harder systems problem than generating a polished answer. Google must keep identity, permissions and source context consistent as work crosses applications and models. A useful enterprise agent must know not only what a user wants, but which records it may read and which actions it may execute.

 

Coworker Agents Get Email, Storage and Audit Trails

Companies can create specialized coworker agents for persistent roles. Each can receive a Workspace account with its own company email address, calendar, Drive storage and directory presence. Google says these agents act under their own identities and can access only the information explicitly shared with them.

 

A legal team could maintain an agent for contract review while finance operates another for reporting. Google says finance and legal configurations are in preview, with versions for government, healthcare and retail planned. The company has not published detailed release dates for those industry packages.

 

Separate identities make activity easier to attribute than when an agent acts invisibly through an employee account. Google says administrators receive audit trails that identify agent actions, while an Agent Sandbox provides a network boundary and an Agent Gateway enforces access policies between agents and company systems.

 

Administrators can also set real-time spending limits. A hard cap pauses an agent rather than allowing an open-ended task to consume unlimited model capacity. That feature addresses a practical risk of autonomous workflows: a badly scoped assignment can multiply calls, subagents and tool activity faster than a person notices.

 

Enterprise Controls Define the Agent’s Real Test

Google’s announcement raises the competitive stakes for Microsoft, Salesforce, ServiceNow and other vendors embedding agents in workplace software. The differentiator is shifting from access to a capable model toward reliable execution across data sources, applications and organizational boundaries.

 

Private preview customers will determine whether the architecture works under real permission structures. Persistent memory can reduce repetitive instructions, but it also increases the need for retention policies, deletion controls and traceable sources. Multi-model routing may improve task fit while complicating data handling and quality assurance.

 

Google has supplied the control framework, not independent proof of broad operational gains. Enterprises will still need measured accuracy, failure rates, latency, cost and human-review requirements for each workflow. Regulated organizations will also need evidence that sandboxing and audit records meet their own legal and security obligations.

 

The launch nevertheless marks a concrete expansion of Google’s enterprise strategy. Gemini is moving from an assistant embedded in individual apps toward a managed worker that can hold context, choose models, delegate tasks and operate with a distinct identity. The private preview will show whether those pieces remain dependable when connected to live business systems.

 

Related Research