Google’s Gemini Agent Spans Apps, Starting With Enterprises

On October 8, 2026, Google Cloud CEO Thomas Kurian introduced Gemini agent at Gemini at Work 2026 as a universal work agent that takes “objectives, not instructions.” It is designed to plan a task, use tools and skills, connect to business systems and return completed work in the applications employees already use. The announcement describes a single agent that can answer questions, create content and write and run code.

Businesses are the initial audience. TechCrunch’s launch account describes work and answers delivered through one interface, with a consumer rollout intended to follow the business launch. For enterprises, the scope of that promise depends on which capabilities their accounts can access and which systems they authorize the agent to use.

What a delegated objective can involve

The proposed workflow starts with an outcome rather than a sequence of prompts. Gemini can break an assignment into steps, choose skills and tools, and produce a result in a document, inbox or developer environment. The design also allows tasks to be scheduled or triggered by an event. Long-running work executes in the cloud, so closing a device need not end the assignment.

For more complex work, Gemini can create temporary specialist agents and coordinate their tasks in sequence or in parallel. A separate coworker-agent mode gives an agent a persistent role, its own organizational identity and storage. That distinction matters operationally: a temporary specialist exists for a particular job, while a coworker can receive changing assignments from a team across sessions.

Business context is intended to reduce repeated briefing. The design combines connected company data, work history, reusable skills and memory of tasks and procedures; teams can also create projects with their own context and tools. Its usefulness therefore rests on the information a company has connected and permitted it to access. A shared prompt box alone does not make every repository or action available.

Where the work can move

Inside Google Workspace, the announced agent is designed to work in Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar while carrying the same context between them. An assignment could therefore draw on a conversation, create a document and continue into a spreadsheet without being briefed anew at each surface. The launch describes these as functions of the same agent, rather than separate assistants for each app.

The wider reach includes Microsoft 365, Slack, Confluence, Jira, Git, Salesforce, ServiceNow, BigQuery, Databricks, Postgres and Snowflake. Model Context Protocol servers and an enterprise tool registry provide routes for additional company tools. These connections define where the agent can obtain context or take an authorized action; the relevant connector and permissions still have to be in place for a particular company.

Model selection is another layer of the design. Gemini can route work among models suited to a task, with Google’s Gemini family and Anthropic’s Claude models named in the announcement and other private and open models described as future options. The agent’s business context and identity are intended to persist when the underlying model changes. Availability of that model choice to customers is narrower than the architectural description suggests.

The controls attached to agent actions

Cross-application work gives permissions a direct consequence: an agent may need to read one system, write in another and run code to finish a task. The announced governance design assigns agents distinct identities, role-based permissions approved by security administrators and audit records attributed to the agent. For an external connection, its identity can be mapped and propagated through authorization standards such as OAuth.

Execution is described as taking place inside an Agent Sandbox with a network boundary. An Agent Gateway applies organizational policy to traffic entering, leaving or passing between agents. A persistent coworker agent is described as seeing only context that team members share with it. Together, those controls determine the authority behind a completed action, including whether a denied request remains denied when work crosses into another application.

Cost controls accompany the same workflow. Smart Routing is intended to select a suitable model for a workload, while a project-level spending cap in the Cloud Billing Console can pause an agent when the cap is reached. Usage accounting includes model tokens and sandbox costs. That makes an action log and a project’s metered spending relevant to the same assignment, particularly when it runs for a long time or delegates steps to other agents.

What is available to enterprises now

Google’s business product FAQ says customers can contact sales or try Gemini Enterprise Plus, but places multi-step background delegation, mobile and desktop access, and third-party model choice in early access. It lists Google Gemini models as the models available today. The announced ability to route work to Claude therefore should not be treated as a model option enabled for every customer account.

The distinction extends to examples from existing customers. Sportswear company On tested dynamic model selection, while other organizations named during the launch use the wider Gemini Enterprise platform. A test of one capability and deployment of the platform establish different things from general availability of the newly introduced agent’s full set of features.

For a company evaluating a specific workflow, the launch claims resolve into four account-level questions:

Access: Is the required delegation mode available in the purchased edition and the organization’s account?
Connections: Are the needed applications and tools connected with the intended read and write permissions?
Authority: Can administrators identify the agent, restrict its actions and attribute its work in audit records?
Operations: Can a long-running task be monitored and paused, including when its project reaches a spending cap?

The next meaningful change for enterprise users will be wider availability of the named early-access capabilities. Until then, the practical scope of cross-app automation is set by each customer’s enabled features, connected systems and delegated permissions.