Google announces Gemini work agents: they keep running after shutdown, but who has access?

Who has access now: announcement versus universal availability

Google announced Gemini agent at Gemini at Work on October 8, 2026. It targets questions, knowledge work, content, and programming: provide a goal and the agent plans and performs needed work.

Google Cloud's SMB article says early access is underway among small and medium-sized business customers, with all-customer access to follow. No universal date, detailed application eligibility, region list, or new-agent price was given.

A Gemini or Workspace account alone therefore does not establish access. Check the organization's actual entry, administrator guidance, and Google access conditions before a pilot. Coming soon does not mean everyone can use it today.

Where does work run after the computer is shut?

The agent uses persistent cloud execution. Google describes hours- or days-long tasks continuing after laptop shutdown, with memory, context, and personalization retained across devices so the same work can be resumed.

The change is execution location: taking a personal computer offline need not interrupt cloud work. It is not unlimited execution or guaranteed completion. Budget caps pause agents, and authorized tools, accessible data, and organization policies still constrain tasks.

Why can one goal create several agents?

Dynamic subagents are a core capability. A main agent temporarily creates separately identified subagents and coordinates parallel or sequential steps. This divides work without requiring the user to orchestrate every step manually.

Google also describes persistent role-based coworker agents, such as a project coordinator, with identity and storage. In Workspace, they can have email, calendar, and Drive and participate as members.

Unexecuted hypothetical example: a small team asks for a product-launch readiness list using shared progress materials, identifying omissions and drafting a summary. Define deliverables, read-only materials, and who confirms results rather than only saying “handle the launch.”

Claude selection does not grant access to all company data

Gemini is the agent's name; the underlying model is another choice. Google says current tasks can be orchestrated across Gemini and Anthropic Claude. Other major proprietary and open models are future plans without a common integration date.

Connections and data require permissions. Each agent has its own identity, granular roles, and audit records. Coworker agents only receive context provided by you or teammates; enterprise collaboration does not mean automatically reading all internal files.

The announcement lists web, mobile, Windows, Mac, Workspace, Microsoft 365, and Slack usage forms. That describes access scope, not simultaneous availability of every channel or connector to every account. Use the organization's actual pilot entry.

What happens at the budget cap, and what costs count?

Google announced project-level AI spending caps in Cloud Billing Console, tracking tokens and sandbox costs. Tokens measure model-processed content, while sandboxes are isolated environments for agent code execution.

At the cap, the project's agents pause and the owner can choose to resume in the console. This gives long tasks a spending stop point, not free allowance or Workspace feature limits counted in pages or video seconds.

Multi-model orchestration and smart routing match models to tasks; they do not establish guaranteed savings for each team. No complete new-agent rate table was published. Confirm organizational settlement before choosing budgets instead of inferring bills from customer cases.

Prepare a small task with a defined endpoint

First confirm actual agent access and organizational eligibility. If unavailable, prepare requirements. Existing Gemini Enterprise features or an event link do not prove approved early access.

Define outputs, permitted materials and tools, and owner acceptance. A first task can draft a checklist or summary from authorized material; verify information, sources, and results before longer cross-system work.

Finally confirm budget, pause indication, and resume entry. Return after closing the computer to inspect the same task's state and outputs, distinguishing running, budget-paused, and complete. These are pilot suggestions, not tasks executed or account access verified here.