Google Cloud on Thursday introduced the Gemini Agent, a universal AI agent for the workplace that the company says can accept a goal, carry out the work in the cloud, and return with the finished result—even if the task takes hours or days.

The launch reflects how workplace AI is moving beyond assistants that answer questions toward software that is assigned work and left to complete it on its own. CEO Thomas Kurian presented the product at Google's Gemini at Work event, saying that "work now starts in the prompt window." Further details are available in Google's official announcement.

The Gemini Agent operates inside Gmail, Docs, Sheets, Slides, Drive, Chat and Calendar, and also reaches into Microsoft 365 and Slack. Google is positioning it as a single AI for work that keeps running in the cloud after a user closes their laptop, meaning a single task can stretch across hours or days.

'Coworker agents' with their own accounts

The most distinctive feature is what Google calls coworker agents. A manager describes a role—events coordinator, for instance—and Gemini creates an agent provisioned with its own Workspace account: an email address, a calendar, Drive storage, and a listing in the company directory. Each agent is tracked in an audit trail separate from any human employee. In effect, the arrangement extends workplace identity—and the accountability that comes with it—to software.

The concept builds on ground already explored by other AI platforms: Grok popularized it with Grok Bot, OpenAI did the same with GPT dots, and Hermes introduced a bots feature offering similar functionality.

Integrations and multi-model orchestration

The new agent connects to Salesforce, ServiceNow, Jira, Snowflake and BigQuery, along with any server that speaks Model Context Protocol—an open standard that functions like a universal plug between AI and software. Nor does it run solely on Google's own models: Gemini currently orchestrates across Google's Gemini and Anthropic's Claude models, routing each job to the model Google considers the best fit for quality and cost.

Safety, identity and containment

Letting software actously raises obvious safety questions, and Google's answer is to treat each agent like an employee. Every agent receives a cryptographically attested identity—a digital ID verified through cryptography—and every action lands in an audit trail attributed to the agent rather than to a person.

All agents run inside an Agent Sandbox, a walled-off environment designed to keep the software isolated from the rest of a system, and all of their traffic passes through Agent Gateway, which Google describes as an "AI network firewall."

Containment has failed before. In July, researchers at Accomplish AI reported that the local mode of Claude Cowork could escape its Linux virtual machine on Macs and reach files on the host computer—a reminder that isolation layers for autonomous software are not guaranteed to hold.

Cost controls

Cost is the other worry, since agents burn through tokens—the chunks of text that AI models read and write, and that customers are billed for. Google says nearly 500 of its customers each processed more than one trillion tokens over the past year. To keep spending in check, the company added real-time spend caps: when a project hits its limit, the agent pauses until someone clicks to resume. The pause-until-resume design addresses a specific risk of long-running work: because agents keep operating after a user walks away, token usage can accumulate without anyone watching.

Availability

Google did not publish pricing or a general availability date for the Gemini Agent. The financial services and legal versions are in preview—two heavily regulated fields where identity and audit features would face close scrutiny—while government, healthcare and retail versions are listed as coming soon. Pricing, a general availability timeline, and how the early previews perform in production are the main open questions as the rollout proceeds.