Google Cloud introduces Gemini agent for delegated work

Google Cloud announced the Gemini agent at its Gemini at Work 2026 event; the cloud.google.com page gives October 8, 2026 as the quoted date.[1][2] Google says the agent can answer questions, handle knowledge work, create images and media, and write and run code.[2] It can plan work, use skills and tools, connect to business systems, and return completed work inside users’ existing work environments, according to Google.[1] Google also says users can assign objectives rather than spell out each step.[2] This makes delegation a central part of the product’s stated design.

Google says users can access the agent from any device through channels including the command line, Google Workspace, Microsoft 365 and Slack.[2] It can also be embedded in third-party apps as a headless agent, without a dedicated user interface.[2] Within Google Workspace, Google says it works directly in Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, bringing its memory, skills and controls into those apps.[2] The agent runs in the cloud and retains one set of memories, context and personalization across devices and channels, according to Google.[2] Teams evaluating it can therefore consider both app-based and embedded use.

Google says assigned work can continue for hours or days after a user closes a laptop.[2] For multi-step work, the agent can create task-specific sub-agents, which Google says can coordinate workflow steps in parallel or in sequence.[2] Google describes three kinds of memory: session memory for an active task, semantic memory built from documents and interactions, and episodic memory that records prior actions.[2] It also says procedural memory includes how work gets done, including skills the agent creates for itself.[2] These are product capabilities described by Google, not independent performance findings.

Google says the agent selects a model for each job, using Gemini and Claude models today, with other private and open models a possibility in the future.[2] Google says matching models to tasks can improve accuracy on difficult jobs and reduce costs on simpler ones.[2] Its stated controls include enterprise security, administration and governance, plus multi-model orchestration, Smart Routing and real-time spending caps.[1][2] The agent was in private preview, with wider availability expected soon for Workspace customers on select Business and Enterprise plans, according to 9to5Google.[3] The evidence does not specify a general release date or price.