Google Cloud announced a workplace Gemini agent at Gemini at Work on October 8, 2026, designed to carry out assignments across business applications rather than simply answer questions. The announcement describes an agent that can coordinate work and connect to company systems, with access through Gemini Enterprise. (Google)

Access remains limited. Google’s product FAQ identifies multi-step background delegation, mobile and desktop access, and third-party model choice as early-access capabilities. Its separate small-business announcement says the agent is in early access with those customers, with broader availability to follow, but gives no dated general rollout. (Google Cloud)

For enterprise buyers, that distinction separates a product announcement from a deployment commitment. The available materials establish a proposed way to assign work across applications, not that every existing Gemini Enterprise customer can enable every feature. They also do not establish a complete price for the newly announced experience or independently verified performance. (Google Cloud)

Which applications are covered

Announcement coverage names Gmail, Drive, Docs, Sheets and Calendar, alongside Microsoft 365 and Slack. The Verge reports that users will be able to chat with the agent and assign tasks from a single Gemini Enterprise interface. Those are descriptions of the announced experience, rather than a verified inventory of operations available in each application. (The Verge)

Google also describes connections to Salesforce, ServiceNow, Jira, Git, BigQuery, Snowflake and Databricks, as well as Model Context Protocol servers. That expands the stated scope beyond office documents to business systems, development tools and data platforms. (Google Cloud)

The procurement question is therefore more specific than whether a familiar application appears on a list. Reading a record, preparing a draft and changing a record are different levels of responsibility. A connector name alone is not sufficient evidence that all three operations are available, included in the same edition or governed by identical permissions.

Background work and delegated assignments

Google describes cloud execution that can continue across devices and support assignments lasting hours or days. It also says the agent can create temporary, task-specific subagents to divide work into parallel or sequential steps. These are announced operating capabilities, not measured completion times or evidence that delegation reliably improves results. (Google Cloud)

The intended benefit is continuity: a worker could hand off an assignment rather than repeatedly reconstruct it in separate conversations. Evaluating that proposition requires more than checking the final document. A useful assessment would examine whether the agent preserves the right context, recognizes incomplete work and avoids treating an intermediate result as an approved outcome.

Model choice also needs qualification. Google’s small-business announcement names Gemini and Anthropic’s Claude, while its product FAQ lists Gemini models as available today and additional models as coming soon. The FAQ explicitly places third-party model choice in early access; the announcement should not be read as universal customer access to Claude. (Google Cloud)

Permissions are not the same as approval

Google describes a coworker configuration with a dedicated identity and email address, receiving the context that team members provide. Its governance announcement includes enterprise-managed permissions, agent-attributed audit trails, isolated execution and an Agent Gateway intended to enforce communication policies. These are vendor descriptions of safeguards, not independent security findings. (Google Cloud)

The materials reviewed do not settle which consequential actions require a person’s approval. An audit trail records activity; it does not, by itself, establish that a human must authorize an action before execution. Similarly, an agent’s permission to access information does not answer whether it may send, modify or share that information without review. (Google Cloud)

For buyers, the unresolved issue is where supervision sits between assignment and completion. Before delegating externally visible work, an organization would need a clear account of approval boundaries, responsibility for reviewing output and the means of stopping an assignment. Those are evaluation requirements, not controls verified here for the new Gemini agent.

A wider contest over agent oversight

The emphasis on governance has a clear industry parallel. Microsoft introduced Agent 365 in November 2025 as a management layer for agents built on Microsoft, open-source and third-party platforms. Its announcement describes an agent registry, access controls, monitoring and security integration—placing administrative oversight alongside the ability to automate work. (Microsoft)

The comparison is about operating priorities, not equivalent functionality or a performance ranking. For IT leaders, it suggests a useful distinction: selecting an agent that can perform a task is separate from establishing how the organization will observe, constrain and account for its actions.

For now, Google’s announcement is best understood as an early-access workplace-agent offering, not a completed procurement specification. Feature-level access is qualified, while complete pricing and action-level human-approval requirements remain insufficiently established in the reviewed evidence. Those gaps prevent a firm assessment of deployment readiness or total cost. (Google Cloud)