Google Cloud says its new Gemini agent can carry the same memory and context across Workspace, Microsoft 365, Slack, and developer environments, while routing jobs to either Gemini or Anthropic’s Claude models. For enterprise users, the promise is continuity: work can move between applications, run after a laptop closes, and involve multiple agents without starting a new assistant session for each step.
Google’s October 8, 2026 Gemini agent announcement describes a single agent and API for answering questions, completing knowledge work, creating images and media, and writing and running code. It can accept scheduled assignments or respond to events without waiting for someone to send another prompt.
Google wants to go beyond an assistant embedded in office software and provide the coordinating layer between employees, enterprise systems, execution environments, and competing model providers.
Availability needs careful reading. Google presents the universal agent as its overarching architecture and explicitly labels Financial Services and Legal specializations as previews. The launch post does not supply a complete per-feature rollout schedule, so it should not be read as confirmation that every capability is enabled for every Gemini Enterprise customer.
The Agent Keeps Working After the Conversation Ends
Google describes Gemini as a cloud-running agent with persistent memory, context, and personalization across devices and channels. Its stated access points include web and mobile, Windows and Mac desktops, command-line interfaces, Workspace, Microsoft 365, and Slack. It can also operate without a dedicated interface as a headless agent inside another application.
Access and execution are separate promises here. Opening the same assistant in several places is useful; Google also says work lasting hours or days can continue after the user disconnects. Scheduled tasks need no one to initiate each run, and events can start work when something changes. Together, those capabilities move the agent closer to workflow automation than conventional chat.
The same memory, skills, and controls are meant to follow Gemini into Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar. That could reduce repeated briefing as a task moves between documents, messages, and analysis.
Persistent context remains a vendor-described capability, not evidence that the agent will always recall the right information. Enterprises will need to evaluate whether it uses relevant context accurately and respects access boundaries as projects and responsibilities change.
Temporary Subagents and Persistent Coworkers Serve Different Jobs
The announcement distinguishes two forms of delegated work.
Temporary subagents are created for a particular job. Google says Gemini can give them individual identities and coordinate parallel or sequential steps, including workflows that continue for hours or days. They divide a larger objective into work that several agents can handle.
Coworker agents have a more durable role: persistent team members with dedicated identities, email addresses, storage, and access limited to the context supplied by users or team members. Google says they can maintain an operational presence across sessions and changing responsibilities.
Administrators will need to account for that difference. A short-lived research subagent and an ongoing departmental agent should not necessarily inherit the same permissions or lifecycle. A persistent agent needs an owner, a defined remit, and a way to handle reassignment.
Google’s existing Gemini Enterprise release notes illustrate that lifecycle concern. A September 21 update permits administrators to transfer ownership of shared agents, but disables their scheduled or event-triggered work until the new owner re-enables it. This documents an existing ownership mechanism; it does not confirm that every new coworker-agent feature follows the same process.
Connections Determine What the Agent Can Actually Finish
Google lists connections to collaboration software, development tools, enterprise platforms, databases, and desktop files. Named systems include Workspace, Microsoft Office and Teams, Slack, Confluence, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, and Snowflake.
Gemini can also work with Model Context Protocol (MCP) servers inside or outside an organization’s network, according to Google. An enterprise tools registry lets teams publish tools for others in the company to use.
Alongside those connections are skills: reusable instructions, knowledge, or workflows packaged into modular prompts. Teams can publish shared skills, while individuals can create personal ones. The intended separation is useful. Tools provide access to systems; skills describe how a particular job should be performed.
A connector name alone does not establish the scope of a finished workflow. Reading a document, querying a database, sending a message, and modifying a record are different operations. Buyers should check the supported actions, permissions, and release status for each connection they intend to use.
The release notes offer concrete examples. Microsoft Outlook actions including replying, forwarding, and moving messages were marked generally available on September 29. An October 2 update marked federated query mode for selected Data Cloud connectors as preview. That mode queries data in place through MCP using each user’s credentials, without ingesting it into a separate data store.
These documented capabilities do not establish that every listed integration supports every action.
Anthropic Claude Sits Beneath Google’s Agent Layer
The model-choice architecture is one of the announcement’s most consequential details: Gemini names the agent, while the underlying model is a separate choice.
Google says the agent can orchestrate work across Gemini-family models and Anthropic Claude models today. Support for other private and open models is described as a future addition.
An enterprise could use different models for different jobs without rebuilding the surrounding context, skills, and tool connections each time. Google argues that matching models to tasks can improve quality on difficult work and reduce costs on simpler work, though the announcement does not establish those gains through a controlled comparison.
Important purchasing questions remain unresolved. The post does not give a comprehensive routing policy, model-by-model pricing schedule, or independently verified assessment of the agent’s selection decisions. “Chooses the best model” is a product claim, not a demonstrated guarantee.
Google is competing for the enterprise agent layer without requiring every task to remain within its own model family. That places it against Microsoft, OpenAI, and Anthropic over who coordinates workplace AI, even while Anthropic supplies models that Google’s agent can use.
Model flexibility does not automatically make the whole agent portable. Memory, skills, identity, and execution can remain tied to the coordinating platform even when the model underneath changes.
More Autonomy Requires More Specific Controls
Google describes per-agent identities, role-based permissions, audit trails, sandboxing, an Agent Gateway, Smart Routing, and real-time project spend caps.
These controls address different risks. Identity and permissions establish which agent can access which resources. Sandboxing limits execution environments, and audit trails help administrators reconstruct activity. Routing and spend caps address the cost of work that may involve several models and subagents.
Sources
- Gemini agent announcementcloud.google.com
- Gemini Enterprise release notesdocs.cloud.google.com
- Agent Platform release notesdocs.cloud.google.com
