Quick answer: Google Cloud announced the Gemini agent as a universal enterprise agent that can answer questions, complete knowledge work, create media, write and run code, and continue long-running jobs in the cloud. It is not a general consumer Gemini rollout: launch coverage says it is currently in private preview, with broader availability planned for selected Google Workspace Business and Enterprise plans.[1][3]
Access-status check
- Available to everyone? No. The launch-day status is private preview.
- Where will it appear? Gemini Enterprise, Google Workspace and connected third-party channels.
- Can individuals switch it on? Google’s announcement does not provide a public self-service activation link.
- Published standalone price? The announcement does not list one.
- Best action now: ask your Google Cloud or Workspace representative about preview eligibility, then prepare a narrow, read-first pilot rather than granting broad access.
Do not confuse this with the normal Gemini app or existing Workspace side-panel features. The new product is designed as a persistent enterprise agent with tools, memory, identities, policies and audit trails.[2][3]
What Google announced
Google describes Gemini as one agent and one API for conversational questions, autonomous objectives and code generation. Instead of giving it every step, a user can assign an outcome; Gemini can plan the work, choose tools and skills, connect to business systems, and return completed work inside documents, inboxes or developer environments.[2]
The announcement positions the Gemini agent for organizations, not as a replacement name for every Gemini product. It can operate as a personal assistant, support a team, or take on a defined organizational role such as an analyst or project manager.[2]
Gemini agent access and availability
| Question | Launch-day answer |
|---|---|
| Current release stage | Private preview, according to 9to5Google’s launch report |
| Broader rollout | Planned for selected Workspace Business and Enterprise plans |
| Consumer Gemini accounts | No general consumer rollout was announced |
| Self-service signup | No public activation workflow appears in Google’s launch announcement |
| Standalone price | No price is listed in the launch announcement |
Availability can change quickly after a private preview. Treat plan lists, regional access and pricing as unconfirmed until they appear in your Admin console, contract or an official Google product page.[3]
Where the Gemini agent works
Google says the agent is designed to operate across web, iOS, Android, Windows and macOS, as well as command-line tools, Google Workspace, Microsoft 365 and Slack. An API can also expose it inside third-party applications without a dedicated user interface.[2]
Inside Google Workspace, Gemini is intended to work directly in Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar. Google outlines three patterns:
- Personal assistance: use the user’s permitted context to coordinate work across apps.
- Proactive delegation: identify a task in an email or document and offer to complete it.
- Coworker agent: create a persistent team agent with its own Workspace account, email address, calendar, Drive storage and directory presence.
A coworker agent acts under its own identity and is supposed to see only the information shared with it. Team members can add it to Chat spaces or mention it in collaborative work.[2]
How it differs from a normal chatbot
| Capability | Practical meaning |
|---|---|
| Persistent execution | Cloud jobs can continue after the user closes a laptop |
| Multi-agent orchestration | Gemini can create temporary sub-agents for parallel or sequential steps |
| Persistent coworker identity | A team agent can have its own email, storage and permissions |
| Tools and MCP | It can connect to approved business software and Model Context Protocol servers |
| Skills registry | Teams can publish reusable workflows and instructions |
| Model routing | The agent can choose among Gemini and supported third-party models for a task |
| Four memory types | Google describes session, semantic, procedural and episodic memory |
Google says current model choices include its Gemini models and Anthropic’s Claude models, with additional private and open models planned. That makes “Gemini” the agent layer while the underlying model can vary by job.[2]
Systems and tools it can connect to
The announced connectors include collaboration and productivity systems such as Confluence, Microsoft Office, Teams, Slack and Workspace; development systems such as Git and Jira; enterprise platforms such as Salesforce and ServiceNow; and data systems including BigQuery, Databricks, Postgres and Snowflake. Google also says Gemini can use approved MCP servers inside or outside a company network.[2]
That breadth is useful, but it also creates a large permission surface. A pilot should connect only the minimum systems required for one measurable workflow.
Security and governance controls Google describes
- Separate agent identity: each agent can receive its own cryptographically attested identity.
- Role-based permissions: administrators can grant fine-grained access rather than letting an agent inherit everything.
- OAuth propagation: identity can be mapped when the agent connects to external systems.
- Audit trail: actions are attributed to the agent, including code run in an agent-created virtual machine.
- Agent Sandbox: tasks run inside a sandbox with a network boundary.
- Agent Gateway: traffic passes through a policy-enforcement layer that Google describes as an AI network firewall.
- Spend caps: administrators can set project-level limits that pause an agent when the cap is reached.
These are product controls, not proof that every deployment is safe by default. Administrators still need to design permissions, data classifications, approval gates, retention rules and incident handling for their own environment.[2]
12-step pilot checklist for administrators
- Confirm eligibility: get written confirmation of the preview, plan, region, data residency and billing terms.
- Choose one workflow: start with a repetitive, reversible task that has a measurable baseline.
- Name an owner: assign one business owner and one technical/security owner.
- Use a separate agent identity: do not run a shared team workflow under an executive or administrator account.
- Apply least privilege: permit only the folders, apps, channels and actions required for the pilot.
- Start read-only: test search, summarization and draft generation before allowing edits or outbound actions.
- Block high-impact actions: require human approval for sending email, deleting data, changing permissions, publishing, purchasing and running production code.
- Define sensitive-data boundaries: exclude secrets, payment data, health records, privileged legal material and unnecessary personal information.
- Set cost limits: use a project budget and hard spend cap before long-running tasks are allowed.
- Review logs: confirm that the audit trail identifies the agent, tool, target, action, outcome and approver.
- Test failure cases: use ambiguous requests, revoked permissions, unavailable tools and malicious document content.
- Measure the outcome: compare completion time, correction rate, unauthorized actions and cost against the old process.
Copy-paste pilot prompts
1. Read-only workflow map
Review only the documents in [APPROVED FOLDER]. Do not edit, send, delete, share, move or create anything. Map the current [WORKFLOW] process, list missing information, and identify every system or permission that would be needed to automate it. Separate observed facts from assumptions.
2. Draft with a human approval gate
Using only [APPROVED SOURCES], prepare a draft [DOCUMENT]. Cite the source file for every factual claim. Do not email, publish, share, upload or modify any source. Return the draft and a list of unresolved questions, then stop for human review.
3. Bounded team-agent task
Work only in [APPROVED PROJECT] and use only [APPROVED TOOLS]. Your objective is [MEASURABLE OUTCOME]. Do not change permissions, contact external people, spend money, delete records or run production code. Log each tool used and each item created or changed. If a required action is outside these limits, stop and request approval.
What not to assume yet
- Do not assume the feature is enabled simply because ordinary Gemini appears in Workspace.
- Do not assume every announced channel or connector is available in the private preview.
- Do not assume all Workspace plans will receive the same capabilities.
- Do not assume an agent’s memory should contain every document it can technically access.
- Do not use customer performance examples as guarantees for your organization.
- Do not start with autonomous email, deletion, payments, production deployment or permission changes.
Frequently asked questions
What is the new Google Gemini agent?
It is an enterprise agent announced by Google Cloud that can answer questions, complete knowledge work, create media, write and run code, use business tools and continue long-running work in the cloud.[1][2]
Is the Gemini agent available now?
Launch coverage reports that it is currently in private preview, with wider availability planned for selected Workspace Business and Enterprise plans. Google’s announcement does not provide a public self-service activation link.[3]
Is this the same as the consumer Gemini app?
No. The announcement describes an enterprise agent connected to organizational data, systems, identities, policies and audit controls.
Can it work in Microsoft 365 or Slack?
Google lists Microsoft 365 and Slack among the channels where the agent can operate, alongside Workspace, command-line tools and third-party applications.[2]
Can it use Claude models?
Google says the agent can currently orchestrate across its Gemini model family and Claude models from Anthropic, with more model choices planned.[2]
Does Google list a price?
The launch announcement reviewed for this guide does not list a standalone price for the new universal agent. Confirm contract, seat and usage charges directly before starting a pilot.
Last checked: October 8, 2026 at 23:03 UTC. Private-preview availability, plan eligibility and pricing can change.