Google Cloud introduced a new Gemini agent designed to work with corporate knowledge, applications and workflows at its Gemini at Work 2026 event on October 8. The company describes the product as a single, universal agent for work that can answer questions, handle knowledge tasks, create content and images, and write and run code. Users provide an objective rather than every individual step, allowing the agent to plan the work and return a result inside tools an organization already uses.
The Gemini agent is controlled through one prompt interface. According to Google Cloud, it can break down an assignment, select appropriate skills and tools, connect to business systems and place completed output in a document, inbox or developer environment. That design positions it as software intended to complete an outcome, rather than only as a conversational assistant.
Connections across Workspace and third-party tools
The agent can work directly inside Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar while carrying the same memory, context and company controls across applications. Google also lists connections to Microsoft 365, Slack, Confluence, Salesforce, ServiceNow, Jira, Git, BigQuery, Postgres, Snowflake and other enterprise systems. Support for connecting to Model Context Protocol servers is included in the official announcement.
Google says the agent is designed to be reached through the web, iOS, Android, Windows and macOS. Because execution takes place in the cloud, a long task can continue after a user closes a laptop, while context can remain consistent across devices. The product is currently in private preview for enterprise customers. Google has not announced broad general availability or final pricing.
Sub-agents and a coworker identity
For multi-step assignments, Gemini can create temporary, task-specific sub-agents and coordinate parallel or sequential work. A separate coworker-agent configuration gives the software a persistent role, storage and its own company email address. Google says that kind of agent receives access only to context that users or team members explicitly share with it.
Google's Workspace examples include identifying usual meeting participants from work history, checking calendars, starting an email thread, researching a market, building a financial model in Sheets and turning the results into a presentation. These are manufacturer-provided scenarios, not findings from an independent performance test.
Task-based model selection
The Gemini agent is separate from the model operating underneath it. Google says the system can currently route work across its Gemini family and Anthropic's Claude models, selecting the model that best fits each assignment. The company plans to support additional private and open models in the future. The stated goal is to preserve quality for difficult work while controlling cost on simpler tasks.
For enterprise security, Google highlights identity and policy management, authorization, permission controls, secure sandboxing and network gateways. A coworker agent operating within Workspace acts under its own identity, while access follows an organization's existing sharing and membership rules. Reliability, error rates and operating costs in real deployments will still require broader usage data and independent evaluation.
Competition shifts toward enterprise agents
Reuters placed the announcement within a wider race among technology companies to sell AI systems that can act more autonomously across applications. Google is emphasizing a unified interface, corporate context, multiple model options and governance controls. Specialized versions for financial services and legal work are in preview, while editions for government, healthcare and retail are planned for later.
The announcement signals a major product direction, but Gemini agent is not yet a finished consumer service available to everyone. For businesses, practical questions include which permissions the software receives, how its actions are reviewed and how consistently it performs over tasks lasting hours or days. Those issues should become easier to measure as the private preview expands.
