Newsletter

Google Cloud launched a Gemini proxy on Thursday, which is a single AI designed for work. It can run in the cloud for hours or even days and can execute tasks across applications such as Gmail, Docs, Sheets, Slides, Drive, Chat, and Calendar.

Managers can create "colleague proxies," assigning them independent Workspace email addresses, calendar access, Drive storage space, and entries in the company directory. Each proxy is separately recorded in the audit trail, distinguishing them from any human employees.

Google has not provided pricing information or a formal full launch date; the financial services and legal versions are currently in the preview stage, while the government, healthcare, and retail versions are listed as "coming soon."

Google Cloud launched a general-purpose AI proxy on Thursday for work scenarios. The company stated that this proxy can receive a task, complete it on its own, and then return with the results. Google CEO Thomas Kurian introduced this product at the company's Gemini at Work events, stating that "work now starts with a prompt window."

This product is named Gemini Agent and can run on Gmail, Docs, Sheets, Slides, Drive, Chat, and Calendar. It can also be integrated with Microsoft 365 and Slack. Google states that even when you close your laptop, it will continue to run in the cloud, so a single task may last for several hours or even days.

The most notable feature is what Google calls "Colleague Proxy." Managers simply need to describe a role, such as an event coordinator, and Gemini will create a proxy with a Workspace account: including an email address, calendar, Drive storage space, as well as a record in the company directory.

This new proxy can connect to Salesforce, ServiceNow, Jira, Snowflake, BigQuery, as well as any servers that support Model Context Protocol. Model Context Protocol is an open standard that functions similarly to a universal interface between AI and software. It also doesn't only run on Google's own models: Gemini currently schedules tasks between the Claude models of Gemini and Anthropic, assigning each task to the model that Google deems most suitable in terms of quality and cost.

Forcing software to act autonomously obviously brings security issues, so Google's approach is to treat each agent as an employee. Each agent is given a cryptographically verified identity, which is a number ID confirmed through encryption, and every operation is logged in an audit trail and attributed to that agent, rather than to an individual.

All proxies run within Agent Sandbox, which is an isolated environment designed to keep the software separate from other parts of the system; all traffic also passes through Agent Gateway, which Google describes as a "AI network firewall". However, isolation is not always effective: in July of this year, researchers at Accomplish AI reported that the local mode of Claude Cowork could escape from the Linux virtual machines on Mac and access files on the host computer.

Cost is another concern, as proxies consume token, which refers to the segments of text read and written by the AI model, and customers need to pay for this. Google stated that over the past year, nearly 500 customers have each processed more than 1 trillion token, and a new real-time expenditure limit has been introduced: when the project reaches this limit, the proxy will pause until someone clicks to continue.

Google has not announced the pricing or official full launch date for Gemini Agent. The financial services and legal versions are currently in the preview stage, while the government, medical, and retail versions are listed as upcoming releases.