Google Cloud launches Gemini Agent, an AI agent that can run for days, has its own email and calendar, integrates with Claude, and is managed like an employee, with nearly 500 customers processing more than 1 trillion tokens.

Google Cloud has just introduced Gemini Agent, a versatile AI agent for the workplace that can take in goals, carry out tasks autonomously, and deliver complete results. At the Gemini at Work event, CEO Thomas Kurian emphasized that work now begins in the prompt window. This move shows that Google wants to shift AI from answering questions as an assistant to acting as a true member of an organization.

Colleague agents with their own Workspace accounts

Gemini Agent works in Gmail, Docs, Sheets, Slides, Drive, Chat, and Calendar, and can also connect to Microsoft 365 and Slack. According to Google, the agent continues running on cloud computing infrastructure even after users shut down their computers, allowing tasks to run for hours or even days.

The most notable feature is the colleague-agent model. A manager simply describes a role, such as an event coordinator, and Gemini creates an agent with its own Workspace account, including an email address, work calendar, Drive space, and information in the employee directory. This model is not new: Grok, OpenAI, and Hermes have all rolled out similar features.

In terms of connectivity, the agent integrates with Salesforce, ServiceNow, Jira, Snowflake, BigQuery, and any server that supports the Model Context Protocol (MCP), an open standard that serves as a common interface between AI and software. Notably, the system runs not only on Google’s models but also coordinates work between Gemini and Anthropic’s Claude, assigning each task based on quality and cost.

Managing agents like employees: the challenges of safety and cost

To address the risks of allowing software to operate autonomously, Google manages each agent much like an employee. Each agent has a cryptographically authenticated identity, and every action is recorded in an audit log and attributed directly to the agent rather than to an individual. Agents run in the isolated Agent Sandbox environment, while all data traffic must pass through Agent Gateway, which Google describes as a network firewall for AI.

However, isolation mechanisms have shown weaknesses. In July, researchers at Accomplish AI reported that Claude Cowork’s local mode could escape a Linux virtual machine on a Mac and access the host machine’s files.

Cost is another concern, as agents consume large amounts of tokens—the units of text that models read and generate, and the basis for calculating fees. Google says that over the past year, nearly 500 customers each processed more than 1 trillion tokens. The company has added real-time spending limits: when a project reaches its limit, the agent automatically pauses until someone authorizes it to continue.

Google has not announced pricing or a general release date. Versions for financial and legal services are in the trial phase, while versions for government, healthcare, and retail will come later. The product’s practical appeal will depend on its price and the reliability of its safeguards.