During the Gemini at Work 2026 event, Google Cloud announced Gemini agent, a new AI agent aimed at enterprise customers and presented as a “unified agent for work.” The core idea is for the user to define the goal rather than provide detailed instructions for every step, while the agent analyzes and executes the task across connected applications and services.
Gemini agent is available within the Gemini Enterprise application and Google Workspace, and it can also work with external services. Google Cloud says it can answer questions, complete knowledge work, create media, and write and run code.
From Personal Assistant to Team Member
The agent may handle individual tasks or act on behalf of a group or a specific organizational role. One example presented by Google involves turning a message from a manager requesting the latest project updates into a slide presentation with one click, after Workspace Intelligence recognizes the task as delegable.
In another example, a user can ask it to organize a meeting with a team of regional events managers without entering their names or email addresses. Gemini attempts to identify the people from the chat-space membership and the context of a previous discussion, then reviews their calendars and starts an email thread to coordinate the time, including external participants.
Coordinating Models and Subagents
Gemini agent operates in the cloud, and Google says it selects the most suitable model for each task. Current options include the Gemini and Claude model families, with plans to support other proprietary and open models in the future. According to the announcement, this approach is intended to improve result quality and reduce cost.
The agent also uses subagents to execute multistep tasks and can coordinate workflows involving parallel or sequential stages that may continue for hours or days. This connects it to external tools and services rather than limiting execution to a single application.
Shared Memory and Broad Integration
The platform relies on a shared set of memory, context, and personalization data. The memory types identified by Google include session memory for the current task, semantic memory for building an organized knowledge base, procedural memory related to how work is carried out, and episodic memory that records what the agent has previously completed.
The agent can be invoked through @Gemini in Gmail, Docs, Sheets, Slides, and Chat. Google also mentions channels such as the command line, Slack, and Microsoft 365, and provides an API for embedding it in external applications as an agent without a dedicated interface.
What Changes in Practice?
The announcement moves Gemini from an assistant model that waits for separate commands to a system expected to manage goals and extended workflows across multiple tools. However, the source provides no details about pricing, permission limits, or oversight and review mechanisms—important considerations for enterprises that may delegate access to their data, calendars, and external services.
Gemini agent is currently available in private preview, and Google Cloud says its broad rollout will begin soon for Workspace customers subscribed to specific Business and Enterprise plans.