THE SHORT VERSION
- Google Cloud announced its Gemini agent at the October 8 Gemini at Work event, describing one agent that can plan tasks and return work in existing business tools.Google Cloud 于 10 月 8 日的 Gemini at Work 活动上宣布 Gemini agent，称它能规划任务并在企业日常工具中返回成果。
- Google says the system can use enterprise context, skills, and connected tools, choose among models including Anthropic’s Claude, and run with identity, audit, and spending controls.Google 表示该系统可使用企业上下文、技能与连接工具，在包括 Anthropic Claude 在内的模型中选择，并配有身份、审计和支出控制。
- TechRadar reported that the agent was still in private preview after the announcement; the vendor’s architecture and customer examples are not independent proof of reliability.TechRadar 在活动后的报道中说代理仍处于私有预览；厂商的架构说明与客户案例不能独立证明产品可靠。
A work agent was the headline at Google’s October eventGoogle 在十月的活动上重点介绍了工作代理
At Gemini at Work 2026 on October 8, Google Cloud announced what it calls the Gemini agent: one work agent that can answer questions, handle knowledge work, create media, and write or run code. The announcement says a user can describe an objective in one prompt, after which the agent plans the work, uses tools, connects to business systems, and returns a finished result in a document, inbox, or developer environment.在 10 月 8 日举行的 Gemini at Work 2026 活动上，Google Cloud 宣布了 Gemini agent。公司把它描述为一个可回答问题、处理知识工作、创作媒体、编写或运行代码的工作代理。公告称，用户可以在一个提示窗口中描述目标；代理随后规划工作、调用工具、连接业务系统，并在文档、收件箱或开发环境中返回完成稿。
The Google Cloud keynote article was published the next day and is adapted from CEO Thomas Kurian’s keynote. It presents the product at architecture scale, including persistent execution, agent identities, model selection, and enterprise controls. That makes it useful for understanding Google’s design goals. It does not mean all of the described features were generally available on the day of the announcement.Google Cloud 的主题演讲文章在次日发布，内容改编自 Google Cloud CEO Thomas Kurian 的演讲。它从整体架构角度介绍了产品，包括持续执行、代理身份、模型选择和企业治理。这些信息能帮助读者理解 Google 的设计目标，却不能说明公告当天列出的每项功能都已经向所有企业开放。
The agent is supposed to plan across more than one app代理计划跨多个应用工作
Google’s proposal goes beyond a chat window that drafts text. The company says the agent can work across systems such as Microsoft Office, Slack, Teams, Salesforce, ServiceNow, Jira, Git, and company databases. It can use a tool registry, reusable skills, and context from an organization’s documents and work history. In Google’s example, a request could move from research to a spreadsheet model and then to a presentation without requiring the user to restate the same project at each step.Google 的方案超出了只会起草文字的聊天窗口。公司称，代理可以跨 Microsoft Office、Slack、Teams、Salesforce、ServiceNow、Jira、Git 和企业数据库等系统工作。它可使用工具注册表、可复用技能，以及来自组织文档和工作历史的上下文。在 Google 的示例里，一个请求可以从研究继续到表格建模，再生成演示文稿，用户不必每一步都重新介绍同一项目。
The product claim is that a person can delegate an outcome and return to completed work. That is a more demanding promise than summarizing a document: the agent must find the right sources, choose a sequence of actions, use the correct account permissions, and produce something the team can inspect. Google says the agent can run scheduled tasks and react to events as well as respond to a prompt. Those are announced capabilities, not a guarantee that every multi-app workflow will complete correctly.产品承诺是用户委派一项结果，然后回来查看完成稿。这比总结一份文档要求更多：代理必须找到正确来源、选择操作顺序、使用正确账户权限，并产出团队能检查的成果。Google 表示 Gemini 除了响应提示，还可以安排任务和对事件做出反应。这些属于宣布的功能，不代表每种跨应用流程都会正确完成。
A persistent agent raises an identity question持久运行的代理会带来身份问题
Google describes the agent as cloud-based, with memory and context available across devices and channels. The company also describes coworker agents that can have a persistent role, a dedicated identity, and their own company email, calendar, and storage. Google says these agents only see the context a person or team shares with them. The keynote describes several kinds of memory: the active task, a structured store built from material the agent reads, procedures for recurring work, and records of previous activity. Google says agent identities are stamped into work logs, while an external connection maps that identity through standards such as OAuth. These design details turn memory and identity into items administrators need to inspect. Google also says Workspace users can delegate from an email and give a coworker agent its own account, then inspect edits in file history. Those are separate product examples, and Google says the agent acts under its own identity. The design is intended to let a team interact with an agent much as it would with a colleague assigned to a role.Google 将代理描述为在云端运行，并能在不同设备和工作渠道间保留记忆和上下文。公司还介绍了具有固定职责和独立身份的同事型代理，它可以拥有专属企业邮箱、日历和存储空间。Google 表示，这些代理只会看到个人或团队分享给它的上下文。主题演讲还描述了几种记忆：当前任务、代理从材料中建立的结构化知识、重复工作的程序，以及过去执行过的活动记录。Google 称代理身份会写入工作日志，并通过 OAuth 等标准映射到外部连接。这些设计让记忆和身份都成为管理员需要检查的内容。Google 还称，Workspace 用户可以从邮件委派任务，为同事型代理分配独立账户，并在文件历史记录中查看修改。这些是不同的产品示例；Google 表示代理会以自己的身份执行操作。设计目标是让团队能像给一位员工分配岗位那样与代理协作。
TechRadar’s report after the event says some agents can receive their own Gmail account, calendar, and Drive allocation, and can work across Microsoft 365 and Slack. That is an independent account of Google’s announced design, not a review of a generally available product. Before an organization creates a service identity for an agent, it should know who owns that identity, who can invite it into a conversation, how credentials are protected, and what happens when the employee who created it leaves.TechRadar 在活动后报道称，部分代理可获得自己的 Gmail、日历和 Drive 空间，并能跨 Microsoft 365 与 Slack 工作。这是对 Google 发布设计的独立报道，并非对普遍可用产品的评测。组织为代理创建服务身份前，应了解身份由谁持有、谁能邀请它加入对话、凭据如何保护，以及创建者离职后如何处理代理。
Model choice is part of the platform design模型选择也是平台设计的一部分
Google says the Gemini agent can select an underlying model for each task. Its October 9 keynote post specifically says the platform can use Google’s Gemini models and Anthropic’s Claude models, with additional private and open models planned. The idea is to keep a team’s context, tools, and data in place while the model doing the work changes. Google presents automatic model selection as a way to balance quality and cost.Google 表示，Gemini agent 可为每项任务选择底层模型。公司在 10 月 9 日的主题演讲文章中明确说，平台目前能使用自家的 Gemini 模型和 Anthropic Claude 模型，今后还计划支持其他私有与开放模型。公司希望团队的上下文、工具和数据留在原处，而实际执行任务的模型可以变化。Google 把自动模型选择描述成平衡质量与成本的方法。
Availability is a separate question. TechRadar reported on October 9 that Gemini Agents were not generally available and that some customers could enter a private preview. Google’s launch page announces the product but does not provide a public signup or a broad-release date in the section used here. A company should not design a production dependency around a feature it can only see in a keynote. Confirm its current preview status, supported region, and service terms with Google Cloud before planning a rollout.产品是否能使用是另一回事。TechRadar 于 10 月 9 日报道，当时 Gemini Agents 尚未普遍开放，部分客户可以进入私有预览。本文使用的 Google 发布页面宣布了产品，但没有给出公开报名入口或广泛发布日。企业不应把主题演讲中看到的功能直接当作生产系统依赖。规划部署前，应向 Google Cloud 确认目前的预览状态、支持地区和服务条款。
Google puts identity, permissions, logs, and spend caps on the listGoogle 将身份、权限、日志和支出上限列为治理要素
Google’s keynote article describes a governance structure for agents. Each agent would have an identity and role-based permissions. Actions would be written to an audit trail under the agent’s identity. Code would run in an Agent Sandbox, and network traffic would pass through an Agent Gateway that applies organization policies. Google also says administrators can set project spending caps and pause agent work when a cap is reached.Google 的主题演讲文章描述了一套代理治理结构。每个代理都会有身份和基于角色的权限。每项操作会以代理身份写入审计记录。代码会在 Agent Sandbox 中运行，网络流量则经过 Agent Gateway，并由它执行组织策略。Google 还表示，管理员可以设定项目支出上限；达到上限时，代理工作会暂停。
These controls are important because a delegated task can touch records, communications, or code. The launch post describes Google’s architecture and product commitment; it does not show independent testing of every connector or a complete policy matrix for each customer configuration. An IT team evaluating the preview should ask which actions are logged, whether logs include tool inputs and outputs, how permissions map to external services, and what happens to an in-progress task after a spend cap pauses it.这些控制很重要，因为委派任务可能读写记录、通信内容或代码。发布文章描述的是 Google 的架构和产品承诺，并未展示对每种连接器的独立测试，也没有提供适用于每类客户配置的完整策略矩阵。正在评估预览版的 IT 团队应询问：哪些操作会记入日志、日志是否包含工具输入和输出、权限如何传到外部服务，以及支出上限暂停任务后，正在进行的工作会怎样处理。
Customer examples remain company-reported results客户案例中的成效数字仍由公司报告
Google’s article includes customer examples from financial services, retail, telecommunications, and other sectors. It says some organizations are using Gemini Enterprise agents in production and reports results such as shorter document reviews or faster issue resolution. Those examples can help a buyer identify possible tasks, but the figures come from Google’s account of customer deployments. The post does not provide a common measurement method or an independent audit for each result.Google 的文章列出金融服务、零售、电信等行业的客户案例。公司称，一些组织已经在生产环境使用 Gemini Enterprise 代理，并报告了缩短文档审核时间或加快问题处理等结果。这些案例可以帮助采购团队寻找可能的任务，但相关数字来自 Google 对客户部署情况的描述。文章没有为每项结果提供统一的测量方法或独立审计。
Keep those stories separate from evidence about the new Gemini agent itself. A customer may be using Gemini Enterprise features that already existed, a limited preview, or a custom deployment. Google’s keynote combines the new agent with a broader set of Workspace, data, security, and industry products. A specific number in a customer example should be linked to the named workflow, period, and measurement before a reader treats it as a likely result for another company.这些客户故事还应与新 Gemini agent 本身的证据分开。客户可能正在使用原有的 Gemini Enterprise 功能、有限预览或自定义部署。Google 的主题演讲把新代理与 Workspace、数据、安全及行业产品一并介绍。读者要参考某个客户案例中的数字时，应确认对应的具体流程、时间范围和测量方法，再判断它是否可能适用于另一家公司。
A useful preview starts with one bounded task从一个有边界的任务开始预览
A team considering the preview can start with a task that has a clear source, a limited set of tools, and a human review step. For example, the agent might prepare a draft project update from a named folder and a specific meeting thread. The reviewer can check whether the facts came from those sources, whether the correct dates and owners appear, and whether the document is saved where the team expects.考虑申请预览的团队可以先选择一项来源明确、工具范围有限且必须经过人工复核的任务。例如，让代理根据指定文件夹和特定会议串准备项目进展草稿。复核者可以检查事实是否来自这些来源、日期和负责人是否正确，以及文档是否保存到团队预期的位置。
Before expanding the pilot, ask what data the agent can retrieve, which actions it can take, how access is revoked, what is written to the audit trail, and how actual costs are capped. Google’s proposal is broad: one agent, many tools, and multiple models. A first trial should be narrow enough that the team can explain every input and every change it made.扩大试点之前，应先问清代理能检索哪些数据、可以采取哪些操作、如何撤销访问、哪些内容会写入审计日志，以及实际成本怎样设上限。Google 的方案很宽：一个代理、多种工具和多个模型。首次试验的范围应小到团队能解释每个输入与每项修改。
Sources & reporting limits来源与核查范围
Follow Products to find related stories on your next visit.关注产品新闻，下次访问时阅读相关文章。
