{
  "id": "gemini-agent",
  "title": "Gemini Agent",
  "lang": "en",
  "source_lang": "en",
  "summary": "Gemini Agent is Google’s task-execution AI, introduced in the Gemini app in 2025 and also used as the name of an enterprise agent announced in 2026.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "product",
  "aliases": [
    "Gemini agent"
  ],
  "generated_at": "2026-10-14",
  "model": "gpt-6.1-sol",
  "url": "https://aiwiki.page/en/gemini-agent/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/gemini-agent/",
      "json": "https://aiwiki.page/zh/gemini-agent.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Gemini Agent is a Google artificial intelligence system designed to carry out multi-step tasks using connected applications, research tools, and web browsing. Introduced as an experimental feature of the Gemini app on November 18, 2025, it extends the assistant from generating responses to coordinating actions on a user’s behalf, with confirmation before critical actions. Google Cloud subsequently announced an enterprise-oriented Gemini agent on October 8, 2026. The consumer feature and the enterprise announcement share a task-delegation approach but have different product contexts and controls. (blog.google)"
    },
    {
      "heading": "Consumer feature",
      "text": "The original Gemini Agent operates within the Gemini app. Its purpose is to accept a relatively broad objective, divide it into smaller tasks, and use appropriate tools to work toward the requested outcome. At launch, it was powered by Gemini 3 and drew on lessons from Project Mariner, Google’s browser-interaction research project. Google initially made it available on the web to Google AI Ultra subscribers in the United States; its release notes specified users aged 18 and older. These were launch conditions rather than a permanent description of availability. (blog.google)\n\nThe feature combines Gemini’s reasoning and tool-calling capabilities with connected Google applications. Its available resources include Gmail, Calendar, Deep Research, Canvas, and live web browsing. This makes it an AI agent in the operational sense: it can use tools to advance a task rather than merely describe the steps a person could take. (gemini.google.com)\n\nTask execution and applications\n\nGoogle’s published examples illustrate several kinds of delegated work:\n\n- Inbox organization: identifying priorities and preparing email replies for approval.\n- Calendar management: managing calendar entries and adding reminders.\n- Travel arrangements: locating flight information in email, comparing car rentals against a budget and vehicle requirements, and preparing a booking.\n- Research-supported workflows: combining information gathered through browsing or research tools with material from connected applications. (blog.google)\n\nThe car-rental example demonstrates the difference between answering a question and executing a workflow. Instead of only suggesting rental companies, the agent connects information from the user’s email with external options and prepares a transaction. Preparing a booking is distinct from completing a purchase: the system is designed to request confirmation before the latter. (blog.google)\n\nUser control\n\nGemini Agent’s advertised autonomy is qualified by human oversight. Google states that it is designed to seek confirmation before critical actions, including sending messages and making purchases. Users can pause execution or take over a task. The intended arrangement is therefore delegated execution with approval checkpoints, not unrestricted authority over accounts or transactions. (gemini.google.com)"
    },
    {
      "heading": "Enterprise Gemini agent",
      "text": "On October 8, 2026, Google Cloud introduced the Gemini agent as a unified agent for work. The announcement described capabilities spanning question answering, knowledge work, content and media creation, and writing and running code. It connects to business systems and uses skills and tools to pursue assigned objectives. This enterprise use of the name should be distinguished from the experimental consumer feature introduced in 2025. (blog.google)\n\nGoogle described the enterprise architecture as supporting conversational interaction, autonomous assignments, scheduled tasks, and responses to events. It also announced access across multiple devices and work environments, including Google Workspace, Microsoft 365, Slack, and third-party applications, as well as operation without a dedicated user interface. These were capabilities described in the announcement, not evidence that every integration was available to every customer. (cloud.google.com)\n\nThe enterprise offering emphasizes organizational controls alongside automation. Google’s announcement lists identity and policy management, authorization and permissions, secure sandboxing, and network gateways. It also describes model selection and cost controls. These features address the administrative requirements of agents operating within business systems and handling organizational information. (cloud.google.com)"
    },
    {
      "heading": "Related products and terminology",
      "text": "The Gemini name covers models, applications, and agent products; those categories are not interchangeable. Gemini 3 was the model generation powering the consumer Agent at its introduction, whereas Gemini Agent was a tool-enabled feature built around that model’s capabilities. A large language model supplies reasoning and generation capabilities, while the surrounding agent product provides access to tools and applications. (blog.google)\n\nGemini CLI is a separate developer-facing interface. Google listed it among the environments supporting Gemini 3, while its official documentation describes commands for managing local and remote subagents. It is not the same interface as the consumer Gemini Agent feature. (blog.google)\n\nGoogle’s May 19, 2026 release notes also introduced Gemini Spark as a personal agent capable of working on users’ tasks. That later name belongs to the evolving consumer-agent product line, but the cited release notes do not establish that every Gemini Agent capability, access condition, or interface transferred unchanged to Spark. (gemini.google.com)"
    },
    {
      "heading": "Limitations and safety",
      "text": "The consumer feature was explicitly introduced as experimental. Its published demonstrations establish intended uses, not a measured success rate across arbitrary tasks. Likewise, Google’s broader Gemini 3 evaluations concern model capabilities and should not be treated as direct measurements of Gemini Agent’s reliability in real-world booking, messaging, or account-management workflows. (blog.google)\n\nTool-enabled browsing also raises cybersecurity concerns. Google reported increased resistance to prompt injection in Gemini 3 and described classifiers intended to prevent browsing agents from following instructions inconsistent with users’ goals. These are protective measures, rather than guarantees that externally supplied content cannot influence an agent. Confirmation checkpoints and enterprise permission controls address different parts of that risk: approval of consequential actions and restrictions on the operations an agent can perform. (blog.google)"
    }
  ],
  "links": [
    "artificial-intelligence",
    "ai-agent",
    "email",
    "automation",
    "large-language-model",
    "gemini-cli",
    "cybersecurity",
    "prompt-injection"
  ],
  "sources": [
    {
      "title": "Gemini 3 brings upgraded smarts and new capabilities to the Gemini app",
      "url": "https://blog.google/products-and-platforms/products/gemini/gemini-3-gemini-app/"
    },
    {
      "title": "Gemini Apps’ release updates & improvements",
      "url": "https://gemini.google.com/updates"
    },
    {
      "title": "15 examples of what Gemini 3 can do",
      "url": "https://blog.google/products-and-platforms/products/gemini/gemini-3-examples-demos/"
    },
    {
      "title": "Google Cloud introduces the Gemini agent.",
      "url": "https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/gemini-at-work/"
    },
    {
      "title": "Welcome to Gemini at Work 2026: Introducing the Gemini agent",
      "url": "https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026?hl=en"
    },
    {
      "title": "Welcome to Gemini at Work 2026: Introducing the Gemini agent",
      "url": "https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026?trk=public_post_comment-text"
    },
    {
      "title": "A new era of intelligence with Gemini 3",
      "url": "https://blog.google/products-and-platforms/products/gemini/gemini-3/"
    },
    {
      "title": "CLI commands",
      "url": "https://github.com/google-gemini/gemini-cli/blob/main/docs/reference/commands.md"
    },
    {
      "title": "Securing the next generation of browsing",
      "url": "https://ai.google/static/documents/ai-responsibility-update-2026.pdf"
    }
  ],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}