Google's Gemini Agent Gets an Email, a Calendar, and an Identity Crisis 🪪

Google Cloud unveiled a universal AI agent for the workplace Thursday, a product that CEO Thomas Kurian said means "work now starts in the prompt window." The offering, called the Gemini Agent, operates inside Gmail, Docs, Sheets, Slides, Drive, Chat and Calendar and reaches into Microsoft 365 and Slack, continuing to run in the cloud after a user closes their laptop so individual tasks can extend across hours or days.

The most prominent feature is what Google calls coworker agents. A manager defines a role, such as events coordinator, and Gemini generates an agent assigned its own Workspace account, complete with an email address, a calendar, Drive storage and a directory listing. Similar approaches have appeared in other agentic frameworks, including Grok's bots feature, OpenAI's GPT bots and Hermes' bots offering. The new agent connects to Salesforce, ServiceNow, Jira, Snowflake, BigQuery and any server that supports the Model Context Protocol, an open standard functioning as a universal connector between AI and software. It runs on more than Google's own models, with Gemini currently orchestrating across Gemini and Claude models from Anthropic and routing each job to the one Google considers the best fit for quality and cost.

To address the safety questions raised by autonomous software, Google said it treats each agent like an employee. Every agent receives a cryptographically attested identity, a digital ID verified through cryptography, and every action is recorded in an audit trail attributed to the agent rather than a person. All agents operate inside an Agent Sandbox, a walled-off environment meant to keep software away from the rest of a system, and all traffic passes through Agent Gateway, which Google describes as an "AI network firewall." Containment has previously failed: 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.

Cost is the second concern, since agents burn through tokens, the chunks of text AI models read and write that customers are billed for. Google said nearly 500 of its customers each processed more than one trillion tokens over the past year, and it added real-time spend caps that pause an agent when a project hits its limit until a human clicks to resume. Google did not publish pricing or a general availability date for the Gemini Agent. Financial services and legal versions are in preview, while government, healthcare and retail are listed as coming soon.

Separately, OpenAI published 722 math manuscripts on GitHub on Tuesday, all produced by an internal model the company has not released. An OpenAI spokesperson said almost all of the output came from a single prompt handed to a single AI agent.