Google has introduced the Gemini Agent, a new AI assistant designed to perform complex tasks across enterprise workflows. During the Gemini at Work 2026 conference, Thomas Kurian, CEO of Google Cloud, highlighted the agent as part of the company's shift toward a single-agent model, rather than relying on a series of instructions. The Gemini Agent is capable of performing both document research and coding tasks from a single interface and a single API, making it versatile for a range of business needs. The agent operates continuously, either independently or in teams, and can handle tasks that span multiple days. Users set a goal, and the agent organizes the work, selecting its skills and tools, then connecting to existing software or any MCP server. It runs in the cloud and maintains the same memory across different platforms, such as Microsoft 365 and Slack. For multi-step tasks, the agent creates temporary sub-agents, each with its own identity. The agent is not limited to Google's models: it directs tasks to Gemini models or to Claude models from Anthropic, a direct competitor, pending the arrival of other proprietary and open-source models. Business versions of the Gemini Agent will be rolled out by sector, with dedicated skills and connectors. In preview: finance and legal. Announced soon, without a specific date: public sector, health, and retail. In Workspace, the Gemini Agent acts directly in emails and documents, suggesting tasks to delegate. If a manager requests a progress update in the form of a presentation via email, a single click is enough to assign it to Gemini. Teams can also create an agent colleague by describing the expected role. This agent receives its own Workspace account: a dedicated email address in the format @agents.enterprise.com, a calendar and a Drive space, and a listing in the company directory. Employees mention it in Chat or in a comment, and its changes appear under its name in the version history. It only accesses content shared with it. Each action by an agent is logged in an audit log under its name, and all its traffic passes through Agent Gateway, a dedicated network firewall for AI that applies the company's rules. Regarding costs, administrators set a spending limit per project in the Google Cloud billing console; once the threshold is reached, the agent pauses until manually restarted. The announcement does not mention the price or availability schedule for the Gemini Agent. It continues to use the name Gemini Enterprise for customer deployments, without detailing the agent's role in this offering or in the Workspace packages. Google launched Gemini Enterprise in October 2025, which absorbed Agentspace, and is based on a catalog logic: companies create their own agents without code using Agent Designer, or choose from agents designed by Google and its partners. On the consumer side, Google introduced Gemini Agent in November 2025 as an experimental feature of the Gemini application reserved for American AI Ultra subscribers. Nearly 80 percent of Google Cloud customers use its AI products, and nearly 90 percent of Fortune 100 companies use Gemini Enterprise. At this scale, organizations have surpassed the experimental stage and now rely on these solutions for their operations. Google also justifies this approach with costs: the company notes that the price per token has dropped by 98 percent since 2024, but the surge in volumes is inflating AI bills. The automatic model selection must thus reserve the more expensive ones for tasks that justify them. The Gemini Agent can handle complex tasks without requiring human validation at each step. It can gather the necessary information from emails, documents, and spreadsheets related to Google Workspace. Operating in the cloud, the agent is accessible from a PC application, a smartphone, or via a browser. This allows users to start a task with it on one device and then continue it on another without having to explain everything again. Google specifies that Gemini can also work with third-party services such as Slack and Microsoft 365, provided access is available. The agent is not limited to in-house models: Google explains that it assigns each task to the most suitable model among the Gemini family but also the Claude models from Anthropic, and that other models, private or open, will follow. The company sees this as a way to improve the quality of results while reducing costs. Gemini can also call on specialized sub-agents, responsible for managing different stages of the work. Google pushes the idea of an AI colleague far, allowing companies to assign a unique identity to these assistants, including a dedicated email address and a persistent storage space. Google had already moved in this direction in May during the Google I/O 2026 event, with Gemini Spark: a personal agent that operates 24/7 on Google Cloud virtual machines, to which tasks can be assigned by email via a dedicated Gmail address. Spark, however, targets individuals subscribed to Google AI Ultra, while the new agent is aimed at businesses. For now, the agent presented by Google Cloud is reserved for businesses during a private testing phase. Google promises a broader availability soon for Google Workspace customers subscribed to certain Business and Enterprise plans, without providing a date or price. Nothing is announced, however, for personal Gemini accounts.