Google Introduces a Gemini Agent Designed to Work Across Apps for Days
An optional coworker setup gives the agent its own email, calendar and storage. Google has not specified availability, licensing or a separate price.
Loading page…
An optional coworker setup gives the agent its own email, calendar and storage. Google has not specified availability, licensing or a separate price.
Listen to this story
At its October 8, 2026 Gemini at Work event, Google Cloud introduced an agent intended to carry work across business apps for hours or days, retaining context in the cloud and coordinating specialized agents. The approach centers on delegating an objective rather than prompting separate assistants, with scheduled and parallel work among its planned capabilities. These capabilities remain design goals: Google has not published independent reliability benchmarks, a general-availability date, or pricing and licensing details.
The agent’s memory is designed to separate active tasks, accumulated knowledge, learned workflows, and previous work.
A team’s Event Planner Agent could have its own Workspace account, but that setup is optional; it can access only information explicitly shared with it.
Connections include Google Workspace, Microsoft 365, Slack, Salesforce, ServiceNow, Jira, Git, BigQuery, and Snowflake; Claude works alongside Gemini, with more models planned.
Google Cloud announced a universal Gemini agent on October 8, 2026, designed to carry assignments across workplace apps for hours or days. At its Gemini at Work event, the company described an agent that retains context between devices and can operate as a team coworker—with its own email address, calendar and storage.
The proposed workflow starts with an objective, rather than a series of instructions for separate assistants. Google says the agent can research information, create documents, write code and coordinate other agents, finding the systems it needs and returning finished work.
Execution happens in the cloud. Google says employees can start an assignment on one device and resume from another without rebuilding its context. The agent is also designed to respond to schedules or events, handle parallel assignments and assemble temporary groups of specialized agents for complex tasks.
Its memory design separates current tasks, accumulated knowledge, learned workflows and previous work. Google intends these layers to preserve an understanding of company processes and people, reducing the need for employees to explain the same background repeatedly. Those are design goals, not independently demonstrated results.
The agent can serve an individual employee or take on shared responsibilities for a team. In Google's coworker example, an Event Planner Agent receives a Workspace account, email address, calendar, Drive storage and company-directory entry. Employees could mention it in Google Chat or ask it to edit documents.
That dedicated account is optional, not a stated requirement for every agent. Google says coworker agents can access only information team members explicitly share with them, rather than inheriting a human employee's full permissions.
Other safeguards include cryptographically verified agent identities, enterprise-managed permissions and audit logs recording actions. An Agent Gateway applies organization-wide rules to communications with external systems. Code runs in isolated environments, with sandboxing setting boundaries around execution.
Google says the agent will work through Gemini Enterprise and familiar interfaces, including Google Workspace, Microsoft 365 and Slack. It can already use Anthropic's Claude models alongside Gemini; support for additional proprietary and open-weight models is planned.
Connections include Salesforce, ServiceNow, Jira, Git and data platforms such as BigQuery and Snowflake. Enterprises can also connect Model Context Protocol servers, which give agents access to external tools and context. Google has not supplied independent comparative benchmarks showing how reliably the agent completes lengthy assignments across applications.
Loading discussion...
Join the conversation
Explain where convenience should give way to tighter permissions.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.