Google представила Gemini agent как рабочую систему с памятью и контролем затратGoogle introduces Gemini agent as a work system with memory and spending controls

Google Cloud introduced Gemini agent at Gemini at Work on October 8. The company describes a unified work system that answers questions, creates content, writes and runs code, and carries out assignments through connected tools. The news is how these capabilities are organized around persistent business context. This is Google's product announcement, rather than an independent evaluation establishing that one assistant can reliably handle every kind of work.

One architectural distinction is especially useful: the agent and its underlying model are separate choices. Google says the system can orchestrate its Gemini models and Anthropic's Claude models, with other model support described as future work. In that design, changing a model need not mean rebuilding the work context. The Gemini name here consequently refers to more than an answer-generating model; it also covers the surrounding organization of tools and tasks.

The company separates tools, reusable skills and memory. Tools connect the agent to business systems, skills describe ways to complete work, and memory carries context between interactions. Cloud execution is also described as allowing long assignments to continue after a laptop closes. These are capabilities claimed by the manufacturer. They do not imply unrestricted access to every organizational document: the same announcement ties access to identities and permissions.

For a coworker agent, Google describes a separate account, a continuing team presence and access limited to shared context. Actions are to be recorded in an audit trail and attributed to the agent rather than a person. Code execution is described as occurring inside Agent Sandbox, with network traffic passing through Agent Gateway under organizational policies. Those mechanisms make the proposed responsibility model clearer, but naming them is not a substitute for checking their configuration and behavior in a particular environment.

A concrete cost feature is a spending cap per project in the Cloud Billing Console. Google says reaching the cap pauses the agent, with the user able to choose whether work resumes. The announcement also assigns different statuses to industry specializations: Financial Services and Legal are in preview, while others are described as coming soon. The complete feature list should therefore not be read as one package already available on identical terms to every customer.

Why it matters

Enterprise AI is increasingly presented as completed work across data and applications, rather than an isolated chat answer. A useful evaluation should show the result, the evidence it used and the access boundary within which it operated. Organizations also need to understand which identity performed an action and what the work cost. Today's announcement addresses that broader system, but the claim of universality still needs to be assessed against ordinary workflows. A polished demonstration can explain the idea; repeated, checked outcomes establish whether it is useful.