Google Cloud Gemini Agent Enters Private Preview

Google Cloud Gemini agent arrived as the headline of Gemini at Work 2026. On October 8, Google said work can start in a single prompt window: the agent holds an organization's business context, plans tasks, uses skills and tools, connects to customer systems and returns finished output inside documents, inboxes and developer environments people already use.

What happened

Google's blog post described a universal agent that chooses models for the job, carries built-in cost controls and ships with enterprise security, administration and governance. VentureBeat reported the agent entered private preview for a select customer group the same day, with wider availability promised but undated, and said it will be included at no additional charge where Gemini Enterprise is available, including existing Workspace business and enterprise subscriptions.

VentureBeat said the agent is meant to run across Google Workspace, Microsoft 365, Slack, desktop and mobile apps and the command line rather than a separate site. Google Cloud CEO Thomas Kurian's announcement, as relayed there, framed a common interface and API spanning web, iOS, Android, Windows, Mac and those productivity suites. Agents can operate as a personal assistant or as a persistent digital coworker; in the coworker setup, Google intends agents to receive their own Workspace-style identity, including email, calendar, Drive storage and directory presence, so staff can assign work as they would to a colleague.

Model choice is deliberately open. VentureBeat reported the agent already supports Anthropic's Claude alongside Gemini models, with more proprietary and open-weight models planned, and said administrators can control which models agents may use while Smart Routing picks among permitted options on cost and difficulty. Memory is split into session, semantic, procedural and episodic stores. Connectors include Salesforce, ServiceNow, Jira, Git, BigQuery, Snowflake, Databricks and Microsoft Teams, plus Model Context Protocol servers.

Why it matters

Google Cloud Gemini agent is Google's bid to turn enterprise AI from a menu of chat assistants into a single assignment-taker that can run for hours or days. That is the same destination Microsoft is chasing with its Copilot Autopilot agent, OpenAI with Dots agents and Anthropic with deeper Workspace editing, according to VentureBeat's competitive overview. Persistent coworker identities raise sharper governance questions than one-shot chat: who owns the agent's mailbox, what licenses apply, and how outbound tool calls are gated.

Google told VentureBeat that agents get cryptographically attested identities, run complex work in isolated containers, face an Agent Gateway that applies company rules, and can access only files explicitly shared with them. Every action is logged against the agent's identity. Those claims will be tested in regulated industries; Google is already previewing finance and legal skill packs with FactSet, LSEG, NetDocuments and iManage hooks. For earlier Google model context on Called It, see Gemini 4 Argon and OpenAI's GPT-6 intelligent UI rollout.

Security architecture will decide enterprise trust in the Google Cloud Gemini agent as much as demo quality. VentureBeat reported cryptographically attested agent identities, isolated containers for complex jobs, an Agent Gateway enforcing outbound rules, file access limited to explicitly shared Workspace items, and full action logging. Google also told VentureBeat that prompts and documents stay inside Gemini Enterprise when third-party models such as Claude are used, though it did not detail what those providers receive at inference time.

Adoption anecdotes from Gemini Enterprise's existing base — nearly 90% of Fortune 100 companies on the platform, BNP Paribas targeting more than 65,000 employees, SOMPO building more than 10,000 custom agents — describe today's footprint, not yet the new universal agent, VentureBeat cautioned. Specialized finance and legal packs with FactSet, LSEG, NetDocuments and iManage connectors show where Google expects regulated early wins. Until general availability and clearer coworker-seat pricing land, private preview feedback on multi-day workflows will be the real scoreboard.

What's next

Private preview customers will pressure-test long-running workflows, sub-agent teams and third-party model routing before a general release date appears. Pricing clarity is still thin beyond the "included where Gemini Enterprise is available" line — VentureBeat noted Google has not said whether dedicated coworker Workspace accounts incur separate seats or how usage caps apply to multi-day jobs. Watch whether Google Cloud Gemini agent wins default distribution through Workspace seats the way search once did, and whether enterprises actually allow agents to hold their own calendars and inboxes at scale.