Google is unifying Gemini around one agent, but only in private preview

Google Cloud has announced the Gemini agent as one interface for knowledge work, media creation, analysis and coding. A user supplies an objective, the agent plans the steps, selects tools and models, then returns finished work inside a document, inbox or development environment. Google says long tasks can continue in the cloud for hours or days.

One agent is meant to replace a map of separate assistants

Gemini is designed to work on the web, mobile devices, desktops, the command line, Google Workspace, Microsoft 365 and Slack. It can also run headlessly through an API. Temporary subagents receive specific jobs, while a persistent coworker agent can have its own corporate identity, email, calendar and Drive storage.

Google separates the agent from the model underneath it. The system can route work across Gemini models and Anthropic's Claude, with more private and open models planned. The company is also adding tool and skill registries, enterprise connectors and four types of memory.

Google is selling the control layer, not only the best model

Ethan Mollick argued that Google's challenge after Gemini 4 would be what it does with a good model. The Gemini agent announcement points in the same direction: value is moving from a standalone model to the interface that holds context, permissions, tools and work history.

For companies, that distinction matters. If they can swap models under one management layer according to task and cost, vendor dependence moves from the model to identity, connectors, memory and audit. Google has an advantage in that layer through Workspace and Cloud, but it also has to unite products that previously lived in separate places.

A private preview cannot yet carry a universal promise

The Gemini agent is in private preview for selected enterprise customers. Broader access is planned for some Workspace Business and Enterprise plans, but Google has not published a firm date or a standalone public price. The claim that nearly 90% of the Fortune 100 use Gemini Enterprise does not describe adoption of the new universal agent.

The hardest work is not opening another connector. An agent with email, a calendar and permission to act needs a precise identity, an audit trail, a sandbox and policy for every action. Google describes those layers, but production use will have to establish their reliability across long jobs.

Cross application work without lost control will decide the product

The first useful test will not be a chat answer. It will be a task lasting several hours in which the agent moves through email, documents and a business system, produces the result, respects permissions and leaves an audit an administrator can understand. It will matter just as much whether a company can change models without losing memory and established workflows.

If Google can deliver those scenarios outside a restricted preview, one interface may finally convert its distribution advantage into a coherent product. If it cannot, Gemini agent will be another Gemini label over the same collection of separate features.

Lilith's verdict

Google is finally building one reception desk instead of a corridor of counters. Now it must prove the same old maze is not still hiding behind it.