Gemini agent

Thomas Kurian introduced the Gemini agent at Gemini at Work 2026 on 8 October, describing it as a universal agent for work: one interface and one API for answering questions, knowledge work, image and media creation, and writing and running code. The user hands it goals rather than instructions and returns to a finished result.

Kurian presents the model as a choice distinct from the agent: Gemini agent picks between Gemini models and Anthropic's Claude models per task, with more models to follow. Around that sit a tool registry that connects to any MCP server, a skills registry (global, departmental, personal) and four memories: session, semantic, procedural, episodic. The language model becomes one component among memory, skills, tools, connectors and orchestration.

The second thread is organizational identity. Agent colleagues get their own account and storage, an attested identity, minimal rights, an audit log in their own name, a network sandbox, and a mandatory pass through Agent Gateway, which applies the organization's policies. In Google Workspace the agent works inside Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, as a personal assistant, a proactive delegate, or a team member with its own account.

Fred Cavazza, writing on X on 9 October 2026, places this in a four-rung progression ending in persistent synthetic collaborators, and reads Microsoft's Copilot Autopilot as a comparable ambition (not sourced in his post). He also notes what Google has not shown: reliability and profitability at scale, with the most advanced announced capabilities still in restricted access.