OpenAI Launches Cloud Workspace Agents in ChatGPT to Automate Enterprise Workflows

OpenAI在ChatGPT推出云端工作区智能体以自动化企业工作流 OpenAI正通过部署异步、基于云端的ChatGPT工作区智能体，将团队自动化拓展至基础聊天机器人之外，此类智能体可在Slack、Google Workspace和Microsoft 365等企业工具间执行多步骤软件任务。 OpenAI演讲者Christina Huang、Ho Joon Cha和Victoria Chernova介绍了ChatGPT中的工作区智能体——这是一种由Codex驱动的异步智能体，旨在跨Google Workspace、Slack和Jira等平台执行复杂的团队工作流。通过现场演示销售会议准备智能体和集成Slack的软件评审智能体，团队展示了非技术用户如何借助自然语言与持久记忆来构建、测试和共享智能体。会议最后还针对管理控制、自定义GPT迁移、定价模式，以及智能体、技能与Codex之间的关键架构差异提供了指导。 OpenAI 推出基于 Codex 驱动的 ChatGPT Workspace Agents，专为团队跨外部软件的异步、多步骤协作工作流设计，即使关闭设备或浏览器也能在云端自主持续运行。 (1:22) 用户无需编写代码即可通过自然语言对话、预设模板或现有 Custom GPT 快速构建 Agent，并可针对 Google Workspace 及 Microsoft 365 等连接器配置细粒度的只读安全权限。 (6:37) Agent 配备实时预览日志以便在部署前排查推理与工具调用轨迹，同时具备持久化内存机制，可在 1:1 对话中按用户独立隔离或在 Slack 频道中跨成员共享团队上下文。 (22:08) Workspace Agents 目前向商业与教育版用户开放研究预览且在 5 月 6 日前免费使用，之后将转为根据任务执行复杂度与运行时长计费的积分制模型。 (35:37)

OpenAI introduced workspace agents in ChatGPT as cloud-based, Codex-powered autonomous systems built for team-wide, multi-step business workflows. Available in research preview for enterprise and education plans, these systems execute asynchronous background tasks across external enterprise platforms using persistent memory, granular permissions, and modular skills. Workspace Agents Architecture and Enterprise Ecosystem (1:22) Christina Huang distinguishes workspace agents from individual Codex tooling and the developer-focused Agents SDK. While Codex serves individual users working directly beside an agent, and the SDK allows developers to embed agentic capabilities into custom applications, workspace agents operate asynchronously in the cloud to execute long-running team workflows even when user devices are disconnected. Internal OpenAI teams have used them to automate website creation from Google Docs, streamline accounting month-end close procedures, and conduct vendor risk reviews spanning financial stability and sanctions exposure. Autonomous Execution and Contextual Collaboration Demos (3:24) Ho Joon Cha demonstrated two core applications: a daily sales meeting preparation agent and an IT software review agent deployed in Slack. The meeting prep workflow connects to Google Calendar, Google Drive, and Gmail to extract schedule context, synthesize research into pre-templated briefs, and deliver morning email digests without requiring code to set up. Users can view action plans, chain-of-thought logic, and tool traces in real time during preview runs. Skills, Persistent Memory, and Slack Integration (22:08) Workspace agents decouple high-level reasoning from discrete capabilities through the use of skills and memory primitives. Speakers define skills as reusable, codified operational playbooks or scripts that dictate how specific tasks should be completed, while the agent serves as the overarching worker orchestrating those skills alongside connected tools. Persistent memory functions like an integrated file system, storing contact details, operational rules, and past outputs. In personal ChatGPT sessions, memory is strictly isolated per user, whereas deployment into a shared Slack channel allows contextual memory to be shared across that channel's interactions. Governance, Tool Support, and Evolution from Custom GPTs (27:34) Huang positions workspace agents as the evolution of Custom GPTs, shifting from static conversational interfaces to scheduled, multi-step external software execution with human-in-the-loop approvals. Enterprises manage agents via role-based access controls, administrative app whitelisting, and compliance audit APIs. Integrations support Google Workspace and Microsoft 365 services—including SharePoint, Outlook, and Teams—respecting existing enterprise permission boundaries. Cross-surface access spans desktop, web, mobile, and Slack, with Teams execution interfaces and collaborative multi-user editing slated for future rollout. Commercial Rollout and Consumption-Based Pricing (35:37) Workspace agents are offered in research preview for ChatGPT Business, Enterprise, EDU, and ChatGPT for Teachers plans, remaining free to use through May 6. Following this preview window, OpenAI plans to transition usage to a credit-based pricing model where consumption scales based on the duration and computational complexity of agent runs, mirroring the structure used for long-running Codex executions. Workspace agents shift automation from interactive chat sessions toward asynchronous, scheduled enterprise execution. However, features such as collaborative multi-user configuration, automated Custom GPT migration tools, native Microsoft Teams chat execution, and specific credit pricing schedules remain in development or pending general release.

OpenAI 团队在 Build Hour 活动中展示了处于研究预览阶段的 ChatGPT 工作空间智能体（Workspace agents），该功能基于 Codex 云端架构，支持跨多种企业软件执行异步多步骤长任务，旨在为团队协作提供持久记忆与自动化流程。 产品定位与核心生态系统划分 (1:22) 工程团队成员 Christina Huang 指出，工作空间智能体（Workspace agents）专为团队级异步协作打造，具备持久记忆并在云端运行，即使关闭浏览器或电脑也能继续执行长周期任务。目前该功能面向 ChatGPT Business、Enterprise、EDU 以及 ChatGPT for Teachers 方案开放研究预览。在 OpenAI 的智能体生态中，工作空间智能体与面向个人直接交互的 Codex，以及供开发人员在自有产品中内嵌功能的 Agents SDK 形成了明确区分。 作为 Custom GPT 的演进形态，工作空间智能体克服了轻量级流程自动化的局限，增加了跨第三方软件的多步骤编排、定时触发调度以及人工审核闭环。OpenAI 内部已在多个职能落地应用：市场团队利用智能体读取 Google Docs 产品简报并结合代码库构建可用网站；会计团队用于加速和规范月末财务结算；财务团队则使用供应商风险审查智能体，自动分析制裁、财务稳定性与声誉风险并生成结构化报告。 销售会议准备的实操构建与权限控制 (3:24) 解决方案工程师 Ho Joon Cha 演示了销售会议准备智能体的构建过程。用户无需编写代码，只需在 ChatGPT 中输入自然语言提示词说明目标、挂接工具与模版，系统即可生成高层次的可编辑蓝图。智能体配置界面的左侧保留对话交互，右侧集中管理调度规则、工具集成和指令。针对外部数据安全，系统支持细粒度应用权限，例如可将 Google Calendar 限制为只读以防止误改日程，同时配置 Gmail 和 Google Drive 仅能检索特定资料及发送摘要。 在技能（Skills）配置方面，智能体可导入团队现有的标准操作指南，或现场通过自然语言生成新技能。演示中智能体根据 Google Docs 模版自动提取客户背景、会议目标与执行摘要。在预览测试模式下，创建者可实时查看智能体的高层计划、思维链推理与工具调用日志。该流程验证后支持团队共享与二度修改（Remix），例如将后端存储从 Google Drive 无缝切换至 SharePoint，或在简报后追加自动生成幻灯片的下游流程。 Slack 集成与企业级自动化审查流程 (17:56) 针对企业 IT 与采购痛点，演示展示了部署在 Slack 内部的软件审查智能体，该设计参考了 OpenAI 内部抑制重复采购与 SaaS 膨胀的真实场景。在测试中，当员工在 Slack 频道申请使用“Screen Studio”录屏工具时，智能体自主调研该软件功能，并对照内部批准列表与采购规则。系统识别出公司已有具备等同功能的批准软件“Bloom”，但进一步检索发现 Bloom 许可证使用率已达 100%，因此智能体自动在 Slack 线程中回复决策逻辑，并同步创建 Jira 工单提请 IT 人员审批例外或增补授权。 构建者可在配置界面指定智能体监听的具体 Slack 频道、响应关键词或“@”提醒触发方式。该场景突出了智能体的记忆（Memory）机制：它充当持久的上下文存储层，用于记录联系人、历史交付与规则，从而提升多步骤执行的连贯性。同时，管理员控制台提供运行追踪（Trace）界面，完整记录每个工具调用与决策步骤，并支持通过 OpenAI API 导出合规日志。 企业治理架构、记忆隔离与商用策略 (27:34) Christina 与 Ho Joon 强调，智能体应被视为承担专门职能的虚拟团队成员，通过“指令”、“技能”（即标准化作业程序与脚本）及工具组合协同工作。针对数据边界，系统实行严格的记忆隔离策略：在 1 对 1 ChatGPT 会话中每个用户拥有独立的私有文件系统记忆；而部署到 Slack 频道时，记忆则由该频道内的所有成员共享。智能体支持保持私有、链接共享或发布至组织内的统一目录，当前仅支持创建者编辑，多用户协同编辑功能正在开发中。 在生态兼容性上，工作空间智能体除 Google 工具外，全面支持 Microsoft 365 体系（包括 SharePoint、Outlook 和 Teams），数据读写遵循用户现有的微软权限与管理员策略，Teams 交互界面的执行支持亦在推进。管理端则通过基于角色的访问控制（RBAC）、应用白名单及合规 API 保障安全。商业化方面，该功能在 5 月 6 日前保持免费研究预览，之后将转向基于积分（Credit）的计费模式，积分消耗将根据单次任务的复杂度与运行时间动态核算。 工作空间智能体展示了从单轮对话向长周期、多系统后台协同的转变。目前多用户协同编辑与 Microsoft Teams 原生执行界面仍在开发中；5 月 6 日研究预览结束后，企业需评估其基于任务复杂度的积分计费模式对实际部署成本的影响。

## Background 该视频由 OpenAI 于 2026 年 5 月 3 日发布在 YouTube 官方频道，属于其“Build Hour”系列技术与实践分享直播。页面简介显示，本期主题聚焦于如何在 ChatGPT 中构建能够跨多种工具协作的“工作区智能体”（Workspace Agents），系统演示如何将团队业务最佳实践转化为共享智能体，涵盖工具与连接器配置、安全护栏设置、云端测试迭代以及企业落地问答。 ## TL;DR OpenAI 团队系统介绍了 ChatGPT 内由 Codex 驱动的“工作区智能体”（Workspace Agents），其专为跨系统长程复杂任务与团队协作打造，具备云端自主执行、持久化记忆及团队共享能力。主讲人现场演示了两个实战用例：一个是能自动检索日程、查阅历史客户笔记并按标准化模板投递每日简报的“销售会议准备智能体”；另一个是常驻 Slack 频道中、可自动核验技术栈与席位利用率并在必要时自动创建 Jira Ticket 的“软件审批智能体”。此外，活动还深入阐述了通过自然语言构建智能体的四种途径、企业级 RBAC 权限与合规审计机制，并明确了与微软生态的集成规划以及 5 月 6 日之后基于额度（Credit-based）的计费规则。 ## Atmosphere & style 全片节奏明快且极具实操感，三位主讲人在产品定位阐述、界面实机操作与架构答疑之间配合默契、行云流水。演示环节穿插着“电视烹饪秀”直接端出成品的风趣调侃，将原本涉及复杂工具编排与底层思维链推理的硬核技术，转化为人人皆可上手的直观对话体验。整场活动兼具企业级软件讲解的严谨规范与极客共创的轻松氛围，让观众深切感受到 AI 从单点提示词助手蜕变为团队常驻数字员工的协作变革。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=kktBVmjA19A&t=0s) OpenAI 推出在 ChatGPT 内由 Codex 驱动的 Workspace Agents（工作区智能体）研究预览版，面向 Business、Enterprise、EDU 及 Teachers 订阅方案开放，支持跨工具长时异步运行且不依赖本地设备。 - [17:34](https://www.youtube.com/watch?v=kktBVmjA19A&t=1054s) Workspace Agents 全面超越传统 GPTs，不仅支持通过自然语言无代码编排与预览调试，还具备云端定时触发、Slack 频道事件常驻监听及自动提单建单能力。 - [30:45](https://www.youtube.com/watch?v=kktBVmjA19A&t=1845s) 智能体引入了基于持久化文件系统的 Memory（记忆）能力，区分 ChatGPT 个人沙箱与 Slack 共享频道两种隔离维度，能够跨任务沉淀上下文并持续自我迭代。 - [34:37](https://www.youtube.com/watch?v=kktBVmjA19A&t=2077s) 该功能在 2026 年 5 月 6 日前提供完全免费试用，之后将转向基于任务复杂度与运行步骤动态计量的额度计费机制（Credit-based pricing）。 ## Key takeaways - [01:22](https://www.youtube.com/watch?v=kktBVmjA19A&t=82s) Workspace Agents 专为团队协同与长时运行设计，与面向个人的 Codex 及面向开发者的 Agents SDK 形成三层互补生态。 - [33:55](https://www.youtube.com/watch?v=kktBVmjA19A&t=2035s) 即便用户合上笔记本电脑或离线，Workspace Agents 依然依托云端沙箱全天候自主执行任务。 - [05:11](https://www.youtube.com/watch?v=kktBVmjA19A&t=311s) 业务专家与非技术人员完全可以通过自然语言对话与向导（Agent Plan）实现复杂工作流的从零搭建与调试。 - [07:39](https://www.youtube.com/watch?v=kktBVmjA19A&t=459s) 安全治理推崇“最小权限原则”，支持在工具连接端一键禁用写权限（如 Google Calendar、Gmail），杜绝误操作风险。 - [36:05](https://www.youtube.com/watch?v=kktBVmjA19A&t=2165s) “技能”（Skills）是企业操作手册与最佳实践的数字化资产，而“智能体”（Agents）则是按需调用技能与工具执行任务的数字员工。 - [11:21](https://www.youtube.com/watch?v=kktBVmjA19A&t=681s) 预览测试（Preview Mode）提供高可见性的白盒追踪，可实时审查智能体的执行计划、工具调用、文档读取及思维链（Chain of Thought）。 - [16:01](https://www.youtube.com/watch?v=kktBVmjA19A&t=961s) 智能体支持在工作区内部公开共享、复制二创（Remix）与换源微调，大幅提升业务流程资产在团队间的复用率。 - [24:28](https://www.youtube.com/watch?v=kktBVmjA19A&t=1468s) 软件审批案例证明智能体能跨越表单提交，自主执行技术栈查重、授权冗余核算、政策比对并在 Jira 自动创建人工升级工单。 - [27:32](https://www.youtube.com/watch?v=kktBVmjA19A&t=1652s) 构建智能体共有四种路径：对话式生成、开箱模板、导入外部资产，以及未来支持的 Custom GPTs 一键迁移工具。 - [29:53](https://www.youtube.com/watch?v=kktBVmjA19A&t=1793s) 除了 Google 套件与 Slack，智能体同样支持无缝集成 Microsoft 365 生态（SharePoint、Outlook、Teams），且更多交互载体正筹备上线。 - [29:08](https://www.youtube.com/watch?v=kktBVmjA19A&t=1748s) 针对企业合规需求，提供基于角色的访问控制（RBAC）、敏感操作人工确认（Human-in-the-loop）以及支持自动化调用的 Compliance API。 - [32:53](https://www.youtube.com/watch?v=kktBVmjA19A&t=1973s) 高绩效智能体的核心原则是设定高度专业化、垂直具体的业务角色定位，而非尝试打造泛用的通用助理。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=kktBVmjA19A&t=0s) 开场与 Build Hour 概览 活动由来自 OpenAI 产品市场团队的 Victoria 宣布开场，并向大家介绍了共同参与本期分享的两位同事：来自 Workspace Agents 工程团队的 Christina 和来自解决方案工程（Solutions Engineering）团队的 Ho Joon。Victoria 介绍道，Build Hours 系列的核心宗旨是帮助各团队通过切实的案例和即学即用的技巧，从 OpenAI 产品中榨取更多实际价值。本期是一个特别专场，主题聚焦于 ChatGPT 中的“工作区智能体”（Workspace Agents）。该功能目前正以研究预览（Research Preview）的形式面向 ChatGPT Business、ChatGPT Enterprise、ChatGPT EDU 以及 ChatGPT for Teachers 订阅方案推出。 议程规划清晰明快：首先由 Christina 对 Workspace Agents 做概括性介绍与认知校准；紧接着由 Ho Joon 带来两个重头戏演示——第一个是能够自主检查日程、检索背景信息并生成会议摘要的“会议准备智能体”（Meeting Prep Agent），第二个则是能够在 Slack 中处理员工申请并根据既定政策路由后续流程的“软件审批智能体”（Software Review Agent）。随后 Christina 会重新返场，分享快速上手的实践心得、阐明 Agent 与传统 GPTs 的异同，并详细拆解企业管理员控制权限；最后环节是直播实时问答（Q&A），屏幕右侧设有提问窗口供观众随时提交疑问。 ### [01:22](https://www.youtube.com/watch?v=kktBVmjA19A&t=82s) 什么是 Workspace Agents：产品定位与生态版图 进入正题后，Christina 首先对 Workspace Agents 的本质进行了解密。从底层技术来看，Workspace Agents 是在 ChatGPT 内部由 Codex 驱动的智能体系统。它们专为处理跨越多个系统、具有长时运行特征（long-running）的复杂任务而设计，天生拥有访问文件、编写执行代码以及调用工具的能力。这类智能体最令人兴奋的突破在于，它彻底超越了过去“单人输入单条提示词（one person with one prompt）”的单点交互模式：它们能够在后台自主搜集所需上下文持续工作，并能无缝共享到 ChatGPT 协作区或 Slack 工作流中，供整个团队协同使用。此外，它们还具备长期记忆，能随对话指导不断迭代优化。 为了帮大家厘清认知，Christina 进一步梳理了 Workspace Agents 与 OpenAI 旗下其他 Agent 产品的边界定位： - **Workspace Agents**：专为“团队协作”设计，核心场景是协同工作，并具备在云端全天候运行的能力，"even when your computer is closed"（即使你合上电脑离线也能在后台持续执行）。 - **Codex**：专为“个人用户”打造，扮演辅助个人高效交付工作的专属贴身助手。 - **Agents SDK**：专为“开发团队”提供，用于将自定义智能体深度嵌入到自有产品与外部客户体验当中。 在 OpenAI 内部，许多业务部门早已将 Workspace Agents 融入日常运营。例如市场团队构建的智能体可以直接把 Google Docs 中的产品需求简报和代码转化为可上线的网页；会计团队打造的智能体让月末财务关账变得更加迅速且标准一致；财务团队则开发了供应商风险评估智能体，能自主搜寻供应商背景、评估制裁风险暴露、财务健康度及商誉风险，最终输出规范严谨的结构化报告。 ```mermaid flowchart TD accTitle: OpenAI Agent 产品生态定位 accDescr: 对比 Workspace Agents、Codex 与 Agents SDK 的目标受众与核心使用场景 subgraph Ecology["OpenAI Agent 产品生态定位"] direction LR A["Workspace Agents"] -->|"团队协同 / 云端长时运行"| A1["团队共享工作流 (Slack/ChatGPT)"] B["Codex"] -->|"个人效能"| B1["个人专属助手"] C["Agents SDK"] -->|"开发者定制"| C1["嵌入自有产品与客户体验"] end ``` ### [03:17](https://www.youtube.com/watch?v=kktBVmjA19A&t=197s) 实操演示：打造销售会议准备 Agent（Auto） 随后话筒交给了 Ho Joon，开启第一个实战演示。对于日常背负大量外部客户沟通指标的商务与交付同学来说，每天赶场参会往往伴随着海量枯燥的案头准备工作。Ho Joon 为此亲手打造了一个名为“Auto”的会议准备智能体。每天清晨，Auto 都会主动发送一封整理妥当的邮件，里面详述当天日程安排，并自动在 Web 或 Google Drive 中调取相关客户过往记录，生成专门的会议准备简报（Brief）。以往每天晚上为了对齐议程、给协同同事写材料要耗费数小时，如今全部由 Auto 在后台代劳。更重要的是，该智能体可以作为团队资产直接共享，方便团队成员根据各自业务流程进行二次定制。 为了让大家无障碍跟随操作，Ho Joon 指引观众参考官方配套的 Cookbook 文档。该教程详尽记录了搭建步骤，便于观众在直播结束后自行演练复现。同时，为了保护真实客户隐私，本次现场实测全部采用虚拟的模拟客户、日程事件、账户笔记和文档模板资产，完整还原端到端的运行闭环。 - 核心收益：免去手动翻查不同业务系统的机械劳动，每天早晨自动投递当天专属会议简报与背景文档。 - 团队复用：通过工作区共享，团队成员可以基于母版调整定制，统一销售团队的客户沟通准备水准。 - 演练材料：提供标准化 Cookbook 操作指引及整套模拟日程与文档示例资产。 ### [05:11](https://www.youtube.com/watch?v=kktBVmjA19A&t=311s) 在 ChatGPT 中通过自然语言构建 Agent Ho Joon 打开 ChatGPT 界面进入构建流程。Workspace Agents 允许使用者全程使用纯自然语言（Natural Language）完成搭建。用户既可以选择从空白画布起步，也可以利用界面下方提供的现成预制模板。为了证明其简易性，Ho Joon 决定直接从一段结构清晰的原始提示词白手起家。 他在提示词中明确指出：这是一个负责销售会议准备的智能体；列出了它所需调用的各种应用与工具链；提供了现成的输出模板；并最后下达调度指令——在工作日清晨发送邮件，汇总即将到来的客户会议并附带完整简报的跳转链接。收到指令后，ChatGPT 会作为构建向导展开对话，首先生成一份详尽的“智能体实施计划”（Agent Plan）供人审查。确认无误后点击构建，左侧面板保留交互对话流，右侧面板则像自动化工作台一样动态生成工具编排、定时调度逻辑与精细化的 System Instructions（系统指令集）。整个过程完全无需编写任何业务代码，却拥有媲美 Codex 引擎在后台跨系统获取数据和长时执行的能力。 ### [07:39](https://www.youtube.com/watch?v=kktBVmjA19A&t=459s) 工具连接与权限管控：配置 Calendar、Drive 与 Gmail 右侧配置面板生成完初步指令集后，进入了最关键的工具连接与权限收敛（Guardrails）环节。用户可以随时通过左侧对话窗口让 ChatGPT 调整配置，也可以在右侧界面手动微调。Ho Joon 特意现场演示了如何进行细粒度的安全控制：“在 Workspace Agents 中，你有绝对的控制权来锁定智能体到底具备哪些操作权限。” 在连接 Google Calendar 时，智能体本职工作只是核对日历项，完全不需要改动或删除日程，因此 Ho Joon 果断禁用了所有日历的写入与修改权限（Write Actions）。这样即便智能体自主运行，团队也能百分之百确信它绝不会误改任何会议安排。紧接着，Ho Joon 同样连接了 Google Drive 用于检索客户档案，并添加了 Gmail 连接器。在 Gmail 端，业务上确实需要它代发会议提醒邮件，因此保留发送能力，但把其他无关的写操作尽数关停。这种“按需授予最小权限”的机制，为企业级落地提供了坚实的防线。在完成基础通讯工具绑定后，Ho Joon 进一步引导大家进入下一个扩展模块——为智能体装配自定义技能（Skill）。 ```mermaid flowchart TD accTitle: 会议准备智能体工具权限配置与执行流 accDescr: 展示 Auto 智能体对 Google Calendar、Google Drive 及 Gmail 的最小权限调度流程 subgraph Integration["工具连接与权限最小化配置"] direction TB GC["Google Calendar"] -->|"只读权限 / 禁用写入"| Agent["Auto 智能体 (Codex驱动)"] GD["Google Drive"] -->|"读取客户文档与模板"| Agent Agent -->|"生成会议简报与摘要"| Agent Agent -->|"写入/发送邮件 (禁用其他写权限)"| GM["Gmail"] end ``` ### [09:32](https://www.youtube.com/watch?v=kktBVmjA19A&t=572s) 为会议准备 Agent 添加技能（Skills）与文档模板 在配置好基础参数后，主讲人继续通过自然语言为 Agent 注入具体的业务逻辑与“技能（Skills）”，指导它如何精准支持会议准备工作。在 ChatGPT 和 Workspace Agents 的体系中，技能的一大核心优势在于能够直接沉淀并复用企业现有的最佳实践。企业日常在各种工具和平台中固化的成熟业务流程，都可以被打包成技能接入 Agent，从而在整个组织内部规模化推广专业经验，保证工作交付的高度一致性。用户可以直接选用组织内现有的热门技能，也可以像配置 Agent 其他环节一样，直接让 ChatGPT 现场起草一份全新的技能并纳入 Agent 的执行手册（playbook）。 在现场演示中，主讲人让 ChatGPT 自动生成技能定义并检索 Google Docs 中的模板文件。Agent 迅速定位到了所需的模板，并补充了具体的排版规范——包括表格、层级标题与要点列表的格式定义。主讲人特别提到，这样生成的会议简报在格式上会非常精致且一目了然，即便是员工在清晨通勤路上用手机查看，也能瞬间抓住所有核心要点。 ```mermaid flowchart TD accTitle: 会议准备 Agent 技能与模板注入流程 accDescr: 展示从知识沉淀到生成排版规范简报的执行流程 subgraph 技能沉淀与配置 A["企业已有业务实践"] --> B["沉淀为 Skill 或由 ChatGPT 实时生成"] C["Google Docs 简报模板"] --> B end subgraph Agent 执行准备 B --> D["解析排版规范<br/>(表格/标题/要点)"] D --> E["生成高可读性会议简报"] end ``` ### [11:21](https://www.youtube.com/watch?v=kktBVmjA19A&t=681s) 无代码预览测试与推理过程观察 为了展示完整的测试体验，主讲人切换到了一个经过多轮反馈打磨的成熟版 Agent。在正式将其发布并设置定时任务前，测试是不可或缺的环节。主讲人直接在构建对话框中向 ChatGPT 输入测试指令："Can you run a preview with my calendar for tomorrow?"（你能拿我明天的日历跑一次预览测试吗？）。这种全程基于自然语言的无代码交互，彻底打破了传统流程自动化的技术壁垒——业务人员和业务专家（SMEs）无需 IT 或工程团队的支持，就能自主搭建、调试和分发属于自己的工作流。 预览测试启动后，界面清晰呈现出 Agent 的完整运行状态。主讲人强调，预览运行的强大之处在于“白盒化”可见性：用户可以实时观察 Agent 的高层执行计划、每一步调用的工具、正在读取的文档、所使用的技能，以及底层的思维链（Chain of Thought）。虽然普通最终用户在日常静默运行时不需要时刻紧盯着这套界面，但对于构建者来说，这不仅能在首次发布前排查行为偏离，还能随时回溯历史运行记录，是验证 Agent 进入生产环境前的关键保障。 - **业务专家自主赋能**：无需等待工程团队排期开发，业务人员通过自然对话即可完成复杂跨工具工作流的调试与上线。 - **透明的执行追踪**：在预览模式下可完整追踪 Agent 的推理链条、文件访问记录与技能调用细节。 ### [14:15](https://www.youtube.com/watch?v=kktBVmjA19A&t=855s) “电视烹饪秀”时刻：高精美度会议简报与实际产出 为了避免在直播中干等漫长的检索过程，主讲人幽默地引用了“电视烹饪秀”（Cooking Demo）的行业行话——就像厨师刚把生食材送进烤箱，紧接着就从另一个烤箱里端出烤好的现成大餐一样，他直接切到了邮箱中该 Agent 此前真实运行生成的结果。主讲人明天的日程排得极满，需要连续会见三家客户：Blossom Mart、Petal Pay 以及 Nectar Works。 邮箱中的最终交付成果令人惊艳：Agent（主讲人将其命名为“Auto”）全自动检索了 Google Drive 中的客户档案并辅以网络搜索，严格遵照技能中所设定的模板规范输出了一封排版精良的邮件。简报中清晰地划分出管理层综述（Executive Readout）、客户画像快照（Customer Snapshot）以及本次会议目标（Meeting Objective）等板块。主讲人感慨道，这份简报不仅让自己在通勤或碎片时间里轻松备会，还能直接同步给参会同事对齐议程和预期产出，每天能为他省下数小时机械繁琐的案头工作。 ### [16:01](https://www.youtube.com/watch?v=kktBVmjA19A&t=961s) 定时触发、全员共享与二次创作（Remix） 完成测试验证后，Agent 即可进入投产阶段。用户首先可以为自身配置定时触发计划，例如在本案例中将其设置为“工作日每日定时运行”。更重要的是，当这种自动化能力在个人身上展现出极高的提效价值后，可以一键开启全员共享。这意味着工作区内的任何同事不仅能直接使用该 Agent，还可以对其进行“复制与二创”（Duplicate and Remix）。 主讲人指出，这种灵活性正是 Workspace Agents 的核心魅力所在。例如，如果另一个团队使用的是微软的 SharePoint 而非 Google Drive，他们只需在复本中通过 ChatGPT 调整对应的数据源配置；或者在简报生成后，追加一个“自动制作配套汇报 PPT”的后续步骤。在总结会议准备 Agent 时，主讲人梳理了标准的构建五步法： 1. **自然语言立项**：用自然语言提示词描述 Agent 目标与定位。 2. **应用权限配置**：连接对应的企业应用与参数访问权限。 3. **能力强化注入**：关联组织技能（Skills）与长期记忆（Memories）。 4. **预览测试排查**：通过实际日程执行预览，验证推理逻辑与输出格式。 5. **发布共享分发**：设定运行周期，共享给团队复用或二次开发。 ### [17:53](https://www.youtube.com/watch?v=kktBVmjA19A&t=1073s) 第二个演示案例：解决软件蔓延的软件审批 Agent 在告别个人提效维度的会议准备 Agent 后，主讲人开启了第二个重头案例——软件采购审批 Agent（Software Review Agent）。与前者不同，该案例完全聚焦于企业内部跨部门的组织级协同，原型直接复刻自 OpenAI 内部 Slack 频道中正在运行的真实生产级 Agent。在几乎所有企业中，都存在一个高频、紧急且敏感的 Slack 采购渠道，员工每天都在其中申请采购各种软件工具。 以往这类渠道中的自动化体验极其生硬，“更像是一个只知道发链接的传统 Chatbot”，员工扔出申请后，往往只收到一堆指引文档链接，依然要自己去搞清楚合规与采购流程。而由 ChatGPT 驱动的 Workspace Agent 则能主动执行复杂逻辑： - **供应商背景调研**：深入调研员工申请采购的新供应商功能与背景资质。 - **现有技术栈查重**：将目标工具的能力与企业已批准采购的存量软件栈进行比对。 - **资源利用率核算**：综合分析现有同类工具的席位使用率与冗余度。 - **决策指引与工单升级**：向申请人给出明确的使用建议，或在需要人工干预（Human-in-the-loop）时直接在 Jira 中升级建单。 主讲人表示，该 Agent 在 OpenAI 内部不仅帮 IT 团队节省了大量审核时间，更直接遏制了重复采购带来的软件蔓延（Tool Sprawl）与预算浪费。随后，他重新打开空白的 Agent 构建界面，准备将采购部门（Procurement Team）预先沉淀的采购工作流最佳实践技能导入其中。 ### [19:58](https://www.youtube.com/watch?v=kktBVmjA19A&t=1198s) 软件审批 Agent 的构建逻辑与数据准备 Ho Joon 继续演示软件审批 Agent 的搭建过程。与之前的会议准备 Agent 类似，搭建的核心起手式依然是明确 Agent 所需调用的系统与工具，为其赋予特定的技能（Skills），然后让 ChatGPT 自动将这些模块串联编排起来。为了节省时间，Ho Joon 没有完整重走一遍从零生成的等待过程，但他热情鼓励大家尝试将自己最拿手的业务流程和专业知识打包，亲手打造一个既能当助手又能全权接管任务的“AI 团队成员”。 在展示 ChatGPT 生成的执行计划（Agent Plan）之余，Ho Joon 详细介绍了该 Agent 背后依赖的数据结构。在生产环境中，这套机制通常会连接企业的软件资产管理平台并通过自定义的 MCP（Model Context Protocol）工具实现；而这次演示为了简化，将数据导入了 Google Sheets。这份模拟数据完整包含了企业内部所有审批通过的软件名单、各个工具目前的实际使用率与授权配额情况，以及每一款软件的具体功能和适用场景。如此一来，当员工为了某个特定需求发起软件申请时，Agent 便能根据现有技术栈进行严密推理，精准判断推荐的已审批软件是否能满足需求。 ```mermaid flowchart TD accTitle: 软件审批 Agent 数据流与决策流程 accDescr: 演示用户在 Slack 发起软件申请后，Agent 结合预置表格和政策规则进行多步推理，最终给出处理结果。 subgraph Input ["请求输入"] A["Slack 频道消息<br/>(员工申请新软件)"] end subgraph Reasoning ["Agent 推理与上下文分析"] B["查询已审批软件库<br/>(Google Sheets / MCP)"] C["核查功能匹配度与<br/>当前 License 占用率"] D["结合公司 IT 政策与<br/>记忆库 (Memory) 判定"] end subgraph Output ["处置动作与输出"] E["自主批准推荐 /<br/>自服务下发"] F["升级至 IT 审批并<br/>自动创建 Jira Ticket"] end A --> B B --> C C --> D D -->|有可用配额且匹配| E D -->|配额超额或需例外审批| F ``` - **数据底座内容**：包含企业预先核准的软件清单（Approved Software Stack）、当前工具的使用率（Utilization）以及功能特性矩阵（Capabilities）。 - **生产与演示差异**：演示环境采用 Google Sheets 承载模拟数据；生产环境直接打通企业级软件管理平台与自定义 MCP 协议。 --- ### [22:08](https://www.youtube.com/watch?v=kktBVmjA19A&t=1328s) Slack 渠道集成、长期记忆（Memory）与集中审计追溯 随后，Ho Joon 切换到一个已经完全配置好的 Agent 版本，重点演示了与 Slack 的集成能力以及 Agent 的核心架构特性。将 Agent 引入 Slack 时，构建者拥有高度细粒度的控制权：不仅可以指定 Agent 监听的具体频道，还可以配置其触发模式——是仅在被 `@` 提及（Mentioned）时响应，还是自动介入频道中所有相关的业务请求。此外，还能补充自定义指令，例如如何处理特定请求格式、是否强制在消息线程（Thread）中回复等。 Ho Joon 还借此机会补齐了前一个演示未及细说的关键特性——“记忆（Memory）”。记忆功能赋予了 Agent 跨任务的持久上下文能力，它能够自行保存备注、联系人信息、过往输出结果等，使 Agent 随着工作流的持续运转变得越来越敏锐聪明。同时，每个 Agent 在后台都配备了集中式的运行轨迹记录（Run History & Trace），管理者可以随时向下钻取（Drill down）审计每一次自主决策的底层步骤，而且所有审计追踪数据均支持通过官方 API 进行导出，满足合规需求。 - **Slack 集成设置项**：支持设定特定工作频道、响应触发条件（@ 提及触发或意图识别自动触发）、以及回复规范（如必须在 Thread 内跟帖）。 - **Agent 记忆能力（Memory）**：构建持久化上下文（Persistent context），支持自动沉淀备忘录与历史输出，随使用频次迭代工作流表现。 - **全局审计与可观测性**：提供中央审计面板追溯 Agent 的完整思考与执行轨迹（Trace），完全支持通过企业级 API 导出日志。 --- ### [24:28](https://www.youtube.com/watch?v=kktBVmjA19A&t=1468s) 实机运行演示：从申请 Screen Studio 到 Jira 工单升级 在 Slack 现场演示中，Ho Joon 发送了一条希望申请使用屏幕录制工具 "Screen Studio" 的消息。他向观众指出，Agent 的响应速度和传统聊天机器人不同，由于它需要执行真正的后台调研和逻辑推理，界面会先弹出 "Working" 的状态指示器。在这几秒钟里，Agent 正在调研 Screen Studio 的功能、核对公司的合规策略、并交叉检索现有已审批软件库。 为了不让大家干等，Ho Joon 打开了一条此前同样请求的实际交互结果。Agent 回复明确指出：该请求需要转交 IT 部门人工审批，而不能直接自服务批准。Agent 逻辑缜密地列出了背后原因：公司内部其实已经采购了一款能够完美替代该高品质录屏需求的软件——Loom，但目前 Loom 的 License 已经处于超额使用状态（Over-utilized），暂时没有多余授权。因此，Agent 自主决定将其升级为人工审批流程，由 IT 部门决定是追加增购 License，还是特批员工采购 Screen Studio。最精彩的是，Agent 甚至免去了员工或 IT 人员手动建单的麻烦，它直接在 Jira 中自动创建好了一张信息完备的待办任务卡片（Task），高效打通了从提问到工单落地的全流程。 - **业务判定依据**：公司既有已核准软件 Loom 满足高画质录屏需求，但由于席位已满（No available licenses），触发了升级规则。 - **免手动建单**：Agent 在判定无法自服务开通后，自动提取申请人意图并在 Jira 中创建 IT 审查 Ticket，列明理由与上下文。 - **业务价值**：极大缓解了采购与 IT 部门处理员工突发性、高时效软件采购请求的沟通负担。 --- ### [27:32](https://www.youtube.com/watch?v=kktBVmjA19A&t=1652s) 构建 Workspace Agents 的四种途径与 GPTs 的演进 演示告一段落，Christina 接过话筒，提纲挈领地总结了企业构建 Workspace Agents 的四种核心路径。第一种是“对话式构建”，只需像日常聊天一样用自然语言向 ChatGPT 描述团队惯常的业务流程，系统便会自动协助搭建；第二种是“模板起步”，利用官方预置好的带有技能与工具的模板快速启动，再按需个性化定制；第三种是“导入现有工作流”，直接从其他平台导入现成的 Skills 和应用连接，无需另起炉灶；第四种则是“从 Custom GPTs 迁移”，如果团队内部此前已经搭建并使用了大量的 Custom GPTs，现在就可以在 Workspace Agents 中对这些工作流进行测试，官方也将很快推出从 GPTs 一键自动转换至 Workspace Agents 的工具。 Christina 进一步澄清了 Custom GPTs 与 Workspace Agents 的承接演进关系。她坦言，当初推出 GPTs 是 OpenAI 探索轻量化流程自动化的第一步，许多团队也借此建立了内部模板；但彼时无论是底层基础模型还是平台基元能力，都还不足以支撑真正强大且高度可扩展的自动化。而 Workspace Agents 则是 GPTs 的下一代形态——它们真正变成了可在团队间共享、能跨多系统执行复杂多步工作流、支持定时排程触发，并且内嵌审批机制与闭环追踪能力的智能体。 | 构建途径 | 适用场景与特点 | |:--- |:--- | | **对话式构建 (In Conversation)** | 适合从零开始，直接用自然语言描述团队现有的业务流程，由 ChatGPT 自动配置工具与编排。 | | **开箱模板 (Templates)** | 适合追求快速上手的团队，利用已内置 Skills 与 Connectors 的模板快速启动并二次微调。 | | **导入外部工作流 (Import Existing)** | 适合已有技术资产的企业，直接从其他平台导入既有的 Skills 与 Apps，复用现有积累。 | | **GPTs 迁移 (Convert from GPTs)** | 针对现有 Custom GPTs 用户的无缝升级通道，近期将上线全自动转换工具。 | --- ### [29:08](https://www.youtube.com/watch?v=kktBVmjA19A&t=1748s) 权限治理、企业控制台与问答环节开启 最后，Christina 深入剖析了面向企业级客户的治理与权限控制方案。她强调，作为构建者，掌控权始终牢牢握在自己手中——你可以自主决定 Agent 可以触碰哪些内部数据源、拥有哪些工具的调用权限，以及在执行敏感高危操作时是否必须经过人类审核（Human-in-the-loop approvals）。对于 Enterprise 和 EDU（教育版）套餐的客户，系统还提供了基于角色的访问控制（RBAC），管理员能够全局定义哪些员工有权使用特定的 Agent，以及开放哪些第三方集成应用。 此外，平台专门提供了 Compliance API（合规 API），供安全团队随时监控、审计并长期追踪 Agent 的调用和数据流向，确保这类强大的自主代理始终在合规合法的安全围栏（Governance）内稳定运行。随着构建机制与安全体系讲解完毕，Victoria 重新回到主持席，宣布进入本场 Build Hour 的在线 Q&A 互动问答环节。 - **构建者控制权**：自主指定 Agent 的数据访问范围与工具集，并针对敏感操作设置强制审批节点。 - **管理员治理工具**：Enterprise 与 EDU 租户支持通过 RBAC 实施细粒度的用户与应用权限管控。 - **合规审计机制**：借助 Compliance API 实现长期自动化审计与安全合规监管。 ### [29:53](https://www.youtube.com/watch?v=kktBVmjA19A&t=1793s) 微软生态集成与运行载体支持 针对很多团队使用 Microsoft 套件（Microsoft 365）的提问，Ho Joon Cha 明确指出，工作区智能体完全能够与微软体系协同工作。此前演示中选择 Google Drive 主要是因为搭建演示环境更为轻便且个人习惯使然，但在生产实践中，智能体同样可以无缝对接 SharePoint、Outlook 以及 Microsoft Teams 等工具。智能体具体能读取或写入哪些数据，依然严格取决于该应用本身的集成边界、用户个人的微软权限配置以及工作区管理员的安全设定。用户完全可以在微软产品套件上完整复刻演示中的所有工作流。 ```mermaid flowchart TD accTitle: Workspace Agent Integration and Surfaces accDescr: 智能体支持集成的主流工具与交互载体架构 subgraph Tools["企业数据与工具生态"] GSuite["Google 生态 (Drive, Docs, etc.)"] MSSuite["Microsoft 生态 (SharePoint, Outlook, Teams)"] Internal["内部知识库与企业数据"] end subgraph Agent["ChatGPT Workspace Agent"] Core["智能体核心 (Codex 引擎 + 规则上下文)"] end subgraph Surfaces["用户交互与运行载体"] LiveChat["ChatGPT 网页 / 桌面 / 移动端"] Slack["Slack 频道及直接交互 (已上线)"] TeamsSurf["Microsoft Teams 交互界面 (筹备中)"] end Tools --> Agent Agent --> Surfaces ``` 在交互和运行界面方面，目前智能体不仅能在 ChatGPT 内部对话使用，还能直接驻留并运行在 Slack 之中。虽然目前智能体尚未作为交互机器人直接进驻 Microsoft Teams 聊天界面，但 OpenAI 正在积极投资底层服务并拓展生态，很快就会推出对更多第三方协同交互载体的原生支持。 ### [30:45](https://www.youtube.com/watch?v=kktBVmjA19A&t=1845s) 智能体持久化记忆机制（Agent Memory） Christina Huang 深入解析了工作区智能体的记忆特性。记忆功能本质上是一个持久化的文件系统（Persistent File System），智能体能够将长期文件和关键笔记存储在其中，以便在后续的多轮或未来运行中重复调用与参考。记忆内容的生成既支持人工通过指令显式添加（Explicit Prompt），也支持智能体根据工作流运转情况自行选择自动归纳写入。 这一持久化文件系统具有严格的渠道隔离（Channel-specific）与权限隔离属性： - **ChatGPT 独立环境**：当用户在 ChatGPT 中使用或他人复制使用该智能体时，每位用户都拥有完全独立的文件系统和实例记忆，彼此互不干扰。 - **Slack 共享频道**：如果智能体部署在团队共享的 Slack 频道中，该频道对应的就是一套统一的共享记忆库，频道内不同成员发送的消息与触发记录将沉淀在同一个上下文中。 - **随时间持续演进**：记忆库使智能体能够记录用户的偏好模式与历史反馈，随着使用时间的推移变得越来越聪明。 ### [31:42](https://www.youtube.com/watch?v=kktBVmjA19A&t=1902s) 智能体共享机制与企业管理控制 关于智能体的分发与权限，Ho Joon 说明企业内部拥有灵活的可见性控制机制。智能体既可以完全设为私有，用于满足个人的专属工作流；也可以通过链接直接分享给组织内的任何同事；或者统一发布到公司的“工作区智能体目录”（Workspace Agent Directory）中供全员查找。需要强调的是，这些智能体被严格锁定在企业的 ChatGPT Workspace 边界之内，外部成员无法跨组织访问。 在企业级（Enterprise）和教育版（EDU）方案中，管理员拥有全方位的掌控力，能够通过基于角色的访问控制（RBAC）来规定谁有权构建、发布和共享智能体。其他员工可以在组织内复制（Duplicate）并重新组合修改（Remix）这些智能体，以便为不同团队定制变体。目前在编辑权限上采取了保守策略，暂时只允许智能体的创建所有者（Owner）进行配置更改，但 Christina 透露，支持多人协同编辑的协作功能很快就会正式推出。 ### [32:53](https://www.youtube.com/watch?v=kktBVmjA19A&t=1973s) 专业化角色设定与智能体调优最佳实践 针对能否创建不同职能角色智能体的问题，Ho Joon 指出，智能体最有效的工作方式正是为其赋予极其专业化、垂直明确的业务角色，而非打造泛用的万灵丹。目前 OpenAI 内部的每一个业务单元（Business Unit）都在使用各自定制的智能体。构建高绩效智能体的核心，在于为其装配该岗位所需的一切要素：适配的业务应用、特定工具、背景文件、专业技能指令以及操作边界。 > "你可以把它们想象成由你亲自打造的高能力团队成员，你为它们提供走向成功所需的一切工具与行动指引。" 得益于记忆机制，团队可以随着时间推移对智能体持续提供指导与校正（Guided and Corrected）。用户只需要通过自然的 ChatGPT 对话界面进行反馈，就能逐步打磨和修正智能体的行为模式与工作风格，因此团队应当尽早动手构建并开启反馈迭代循环。 ### [33:55](https://www.youtube.com/watch?v=kktBVmjA19A&t=2035s) 云端自主运行与多端访问 在客户端依赖方面，Christina 明确澄清，使用工作区智能体完全不需要安装本地桌面应用。从智能体的创建、调试、共享到最终运行，都可以完全在 Web 端完成，同时也全面适配桌面端和移动端应用。更重要的是，智能体完全运行在云端沙箱中，即使用户合上笔记本电脑或离线，智能体依然能够照常执行任务。 Ho Joon 补充道，当天演示的两个智能体均没有在任何本地硬件上运行，全部依托云端承载： - 销售会前准备智能体（Sales Meeting Prep Agent）：基于定时任务（Schedule）在云端自主运行，在每天清晨通勤前准时将整理好的简报发送到手机上。 - 软件审查智能体（Software Review Agent）：常驻在云端，全天候侦听 Slack 频道中的用户申请并自主进行多工具联动审查。 ### [34:37](https://www.youtube.com/watch?v=kktBVmjA19A&t=2077s) 研究预览期与额度计费规则（Credit-Based Pricing） 针对观众高度关心的资费问题，Ho Joon 宣布工作区智能体目前正处于研究预览阶段（Research Preview），所有用户在 5 月 6 日之前均可**免费体验**。他反复鼓励团队抓紧这一窗口期充分进行高强度测试，将其融入真实生产场景并检验性能极限。 在 5 月 6 日之后，平台将切换为基于额度的计费模式（Credit-based Pricing）： - **动态消耗机制**：单次智能体运行所消耗的点数，将根据任务的复杂度、运行步骤和调用的资源量动态计算。 - **对齐复杂编程任务**：其机制类似于 OpenAI 针对复杂长程任务的底层引擎消耗逻辑，因为这类智能体同样具备执行耗时较长、包含多步骤推理与环境交互的复杂任务能力。 - **后续政策公布**：随着 5 月 6 日的临近，OpenAI 将通过官方管理控制台和账单支持渠道向管理员发布详细的定价格局与计费指引。 ### [35:33](https://www.youtube.com/watch?v=kktBVmjA19A&t=2133s) 工作区智能体与 Codex 编码工具的区别 针对工作区智能体与底层 Codex 技术差异的提问，Christina 从产品定位与运行架构给出了明确界定： | 维度 | 传统 Codex / 开发者工具 | 工作区智能体（Workspace Agents） | |:--- |:--- |:--- | | **协作面向** | 主要面向单兵开发者或代码编写环节 | 原生面向企业跨职能团队（销售、法务、IT 等全员协作） | | **运行宿主** | 常依赖本地环境或特定代码编辑沙盒 | 纯云端独立长程运行，电脑休眠关闭依然持续作业 | | **触发与交互载体** | 命令行或 IDE 触发 | 多渠道部署（Slack 监听、网页触发、定时调度等），未来支持更多触发器 | 智能体实际上是将强大的云端计算和代码执行能力打包成了普通业务部门拿来即用的形态，彻底消除了非技术岗位的技术门槛。 ### [36:05](https://www.youtube.com/watch?v=kktBVmjA19A&t=2165s) “技能”（Skills）与“智能体”（Agents）的协同关系 在问答环节的最后，团队理清了许多用户普遍困惑的概念——“技能”（Skills）与“智能体”（Agents）的边界与关系。Ho Joon 给出了一套非常直观的比喻模型： ```mermaid flowchart TD accTitle: Skills vs Agents Execution Flow accDescr: 技能作为操作手册被智能体调用的关系图 subgraph S["组织资产：技能库 (Skills)"] S1["操作手册 / Playbooks"] S2["业务最佳实践"] S3["自动化脚本 / API 说明"] end subgraph A["执行实体：工作区智能体 (Agent)"] Brain["智能体调度核心 (Worker)"] Memory["持久化记忆 (Memory)"] Instructions["系统总体角色指令"] end subgraph T["外部集成工具 (Apps & Tools)"] T1["Google Drive / SharePoint"] T2["Slack / CRM / ERP"] end S -->|作为专业能力插件装配| Brain Instructions -->|指导行动方向| Brain Memory <--> Brain Brain -->|调用技能并操作工具| T ``` - **技能（Skills）是操作手册与业务规范**：技能是团队最佳实践（Best Practices）、作业规程（Playbooks）、指导脚本或固定操作步骤的数字化沉淀，它定义了“某项专业动作应该如何规范执行”。 - **智能体（Agent）是实际干活的数字员工**：智能体是真正调用技能、操纵应用并推进流程的责任主体（Worker）。一个智能体可以同时配备多种不同技能，并在面临具体业务场景时，动态选择最合适的技能与第三方工具来交付完整结果。 Victoria 在最后进行了总结，提醒参会者查阅官方推出的《销售会议准备智能体构建指南》（Sales Meeting Prep Cookbook）以及配套演示资源，并邀请大家报名接下来聚焦于 API 能力的 Build Hour 直播专场。 ## Notable quotes > [01:22](https://www.youtube.com/watch?v=kktBVmjA19A&t=82s) even when your computer is closed — Christina Huang > [01:22](https://www.youtube.com/watch?v=kktBVmjA19A&t=82s) one person with one prompt — Christina Huang > [07:39](https://www.youtube.com/watch?v=kktBVmjA19A&t=459s) 在 Workspace Agents 中，你有绝对的控制权来锁定智能体到底具备哪些操作权限。 — Ho Joon Cha > [11:21](https://www.youtube.com/watch?v=kktBVmjA19A&t=681s) Can you run a preview with my calendar for tomorrow? — Ho Joon Cha > [17:53](https://www.youtube.com/watch?v=kktBVmjA19A&t=1073s) 更像是一个只知道发链接的传统 Chatbot — Ho Joon Cha > [32:53](https://www.youtube.com/watch?v=kktBVmjA19A&t=1973s) 你可以把它们想象成由你亲自打造的高能力团队成员，你为它们提供走向成功所需的一切工具与行动指引。 — Ho Joon Cha > [36:05](https://www.youtube.com/watch?v=kktBVmjA19A&t=2165s) 技能是团队最佳实践、作业规程、指导脚本或固定操作步骤的数字化沉淀，它定义了‘某项专业动作应该如何规范执行’。 — Ho Joon Cha > [36:05](https://www.youtube.com/watch?v=kktBVmjA19A&t=2165s) 智能体是真正调用技能、操纵应用并推进流程的责任主体。一个智能体可以同时配备多种不同技能，并在面临具体业务场景时，动态选择最合适的技能与第三方工具来交付完整结果。 — Ho Joon Cha ## Mentioned OpenAI, ChatGPT, Workspace Agents, Codex, Agents SDK, Custom GPTs, Victoria Chernova, Christina Huang, Ho Joon Cha, Google Calendar, Google Drive, Google Docs, Google Sheets, Gmail, Slack, Jira, Microsoft 365, SharePoint, Outlook, Microsoft Teams, Loom, Screen Studio, MCP, Compliance API, Sales Meeting Prep Cookbook ## Video description Build workspace agents in ChatGPT that can do work across your tools and help teams move faster. In this Build Hour, you’ll learn how to turn your team’s best practices into shared agents from start to finish. Christina Huang (Engineering), Ho Joon Cha (Solutions Engineering), and Victoria Chernova (Product Marketing) cover:  • How to build agents for common workflows that require multiple steps, tools, and sources of context  • Using agent builder chat, choosing tools and connectors, and adding guardrails • Testing, iterating, and sharing agents powered by Codex in the cloud • Best practices and live Q&A  👉 Workspace agents in ChatGPT: https://chatgpt.com/agents 👉 Workspace agents blog: https://openai.com/index/introducing-workspace-agents-in-chatgpt/ 👉 How to guide: https://developers.openai.com/cookbook/articles/chatgpt-agents-sales-meeting-prep 👉 Sign up for upcoming live Build Hours: https://webinar.openai.com/buildhours 00:00 Introduction 01:22 Intro to Workspace agents 03:24 Demo: Meeting Prep agent 18:02 Demo: Software Review agent 27:34 Getting Started 29:53 Q&A [Source](https://www.youtube.com/watch?v=kktBVmjA19A)

[1:22] Christina Huang: even when your computer is closed [1:22] Christina Huang: one person with one prompt [7:39] Ho Joon Cha: 在 Workspace Agents 中，你有绝对的控制权来锁定智能体到底具备哪些操作权限。 [11:21] Ho Joon Cha: Can you run a preview with my calendar for tomorrow? [17:53] Ho Joon Cha: 更像是一个只知道发链接的传统 Chatbot [32:53] Ho Joon Cha: 你可以把它们想象成由你亲自打造的高能力团队成员，你为它们提供走向成功所需的一切工具与行动指引。 [36:05] Ho Joon Cha: 技能是团队最佳实践、作业规程、指导脚本或固定操作步骤的数字化沉淀，它定义了‘某项专业动作应该如何规范执行’。 [36:05] Ho Joon Cha: 智能体是真正调用技能、操纵应用并推进流程的责任主体。一个智能体可以同时配备多种不同技能，并在面临具体业务场景时，动态选择最合适的技能与第三方工具来交付完整结果。

[3:24] Ho Joon Cha: The hours I would spend every evening building out my meeting prep docs... that's all done for me behind the scenes. And the great part of this is that this agent can be shared so that others can customize it for their specific workflow.

[0:03] Hello and welcome back to Build Hours. [0:05] I'm Victoria from product marketing and [0:07] I'm here with Christina and Hojun. Hi, [0:09] I'm Christina and I work on the [0:11] engineering team for workspace agents. [0:13] Hi, I'm Hojun. I'm on the solutions [0:15] engineering team. Awesome. And today [0:17] we're going to show you workspace agents [0:18] in ChatGPT. [0:20] So if this is your first Build Hour, um [0:24] Build Hours is all about helping teams [0:25] get more practical value from OpenAI [0:27] products with real examples and tips you [0:29] can use after the session. So today is a [0:32] special session focused on building [0:33] workspace agents in ChatGPT. And [0:36] workspace agents are available in [0:37] research preview today to ChatGPT [0:39] business, enterprise, EDU, and ChatGPT [0:42] for teachers plans. [0:44] So here's what we're going to cover. [0:45] First, Christina is going to start off [0:47] with a quick intro to workspace agents. [0:49] Then Hojun is going to walk through two [0:51] demos. The first is a meeting prep agent [0:53] that checks your calendar, does [0:55] research, and creates a meeting brief. [0:56] The second is a software review agent [0:59] that helps handle employee requests in [1:00] Slack and roots next steps based on your [1:02] policies. After that, Christina will [1:04] come back to share a few tips on getting [1:06] started, how to think about agents and [1:08] GPTs, and go into some detail on [1:10] enterprise admin controls. And then [1:12] we'll wrap up by answering your [1:13] questions live. So on the right side of [1:15] the screen, there's a Q&A chat box where [1:18] you can submit questions at any time [1:19] during the session. [1:21] So now I'll hand it over to Christina. [1:24] Great. So before we get into the demos, [1:25] I wanted to take a minute to to level [1:27] set on what workspace agents actually [1:29] are. So at a high level, workspace [1:31] agents are Codex-powered agents in [1:33] ChatGPT. They're built to handle [1:35] complex, long-running work that spans [1:37] multiple systems and they have access to [1:39] files, code, and tools. And so what [1:41] makes this really exciting is that [1:43] they're not just helping one person with [1:44] one prompt, but they can actually gather [1:46] the right context, keep going in the [1:48] background, and be shared either in [1:50] ChatGPT or Slack so a whole team can use [1:52] them together. [1:54] And they also have memory so they can be [1:55] guided in conversations and actually [1:57] improve over time. [1:59] So, getting started is quite simple. You [2:00] click agents in the ChatGPT sidebar, [2:02] describe a workflow that your team [2:04] already does, and ChatGPT helps turn [2:06] that into an agent. [2:09] So, today we're focused on workspace [2:10] agents, but since I mentioned Codex, I [2:12] wanted to also quickly explain how [2:14] workspace agents fit in with our other [2:16] agentic products. [2:17] The simplest way to think about it is [2:19] workspace agents are for teams. They're [2:21] built for shared work, for tasks that [2:23] run in the cloud even when your computer [2:25] is closed. [2:26] Codex is great for individuals working [2:28] with a personal agent to get work done, [2:30] and the Agents SDK is for teams that [2:31] want to build custom agents directly [2:33] into their own products and customer [2:35] experiences. [2:37] At OpenAI, many teams are already using [2:39] workspace agents across a few different [2:41] functions. [2:42] As a couple of examples, our marketing [2:44] team has built an agent that turns a [2:46] product brief directly into a website, [2:48] and this pulls requirements from Google [2:49] Docs and code. Um our accounting team [2:52] built an agent that helps prepare for [2:53] month-end close faster and more [2:55] consistently, and our finance team has [2:58] built an agent for a vendor risk review, [3:00] which researches vendors, assesses [3:02] signals like sanctions exposure, [3:03] financial health, and reputational risk, [3:05] and produces a structured report. [3:09] So, today Hojin is going to walk through [3:11] building two of these examples, um the [3:13] meeting prep agent and the software [3:14] reviewer. So, I'll hand it over to him. [3:16] Yeah, thanks Christina. So, uh we'll be [3:19] walking through how to build these [3:20] agents, um and we're going to start off [3:22] with this meeting prep agent. Um now, [3:25] I'm sure folks who are customer-facing [3:28] on this call, um you have just a lot of [3:30] uh customer meetings to get to every [3:32] single day. Uh so, this is an agent that [3:35] I built to help with the manual [3:37] multi-system work of customer meeting [3:39] prep. [3:40] Uh so, as you can see on the right side [3:42] here, every single morning I get a email [3:45] from my agent uh named Auto, which [3:47] checks my calendar, uh it does research [3:50] on the customer, whether it's on the web [3:52] or within my Google Drive. And I'll [3:54] actually create a meeting brief for each [3:56] meeting that it will link and provide to [3:58] me. So, you know, the hours I would [4:00] spend every evening building out my [4:02] meeting prep docs, uh docs that I want [4:04] to share with my team members, so uh [4:06] we're aligned on the agenda. That's all [4:08] done for me behind the scenes. Um and [4:10] the great part of this is that this [4:12] agent uh can be shared um so that others [4:15] can customize it for their specific [4:17] workflow um and get the similar value [4:20] that I'm seeing. So, let's jump in. [4:23] Uh what we'll do is actually we'll go [4:24] through this cookbook, uh which um [4:27] actually walks you through how to create [4:29] the agent that I just described, uh so [4:31] you can follow along with what I'm doing [4:33] on screen, or you can also, you know, [4:35] reference this and know that you can go [4:37] back to it um later on after this [4:39] session. [4:41] What I'll also introduce you to are some [4:42] of the demo assets that I'll be using [4:44] here today. Uh don't worry, these are [4:46] just mock events uh with some fake [4:48] customers, but I just wanted to show you [4:50] a real-live test run uh towards the end [4:53] of this demo section. Uh so here we're [4:55] going to go through um my calendar, and [4:58] then we'll also have uh some, you know, [5:00] sample assets uh to show you how it all [5:03] comes together. Uh so some account [5:05] notes, um other company details, um and [5:08] then a sales meeting prep to bring it [5:09] all together. [5:11] So, let's dive into ChatGPT and kick off [5:14] this agent build process. [5:16] Workspace agents um is going to allow [5:19] you to build agents uh just using [5:21] natural language. So, you can start from [5:22] a blank slate, or you can utilize some [5:25] of our friendly templates down below uh [5:28] to get started with your agent. Um but I [5:31] want to show everyone how easy it is to [5:33] build these agents from scratch. So, [5:35] let's get started with a prompt uh that [5:37] just outlines what we discussed uh in [5:39] terms of the use case, and it will get [5:41] me started uh building and working with [5:43] ChatGPT. [5:45] So here, my prompt just calls out that [5:47] it's going to help me with sales meeting [5:49] prep. It calls out the various tools and [5:51] apps that I might need. Um in this case, [5:53] it also says here, I have a template for [5:56] you to use. Uh and then last of all, you [5:58] know, just directing it to send an email [6:00] with a summary of my upcoming meetings [6:03] and links to the full briefs. [6:05] So, if I if I enter in that prompt, uh [6:08] what you'll see is that Chat GPT is [6:09] going to guide you throughout the build [6:11] process. [6:12] Um it's going to create an outline for [6:14] the agent plan that I can review before [6:16] getting started building. Um but I [6:18] really can just work with Chat uh GPT [6:21] here. Um just use natural language, give [6:23] it feedback, iterate together, and we [6:26] can build a really powerful complex [6:28] agent. [6:29] So, here I have um [6:31] you know, the agent plan, everything I [6:33] need, the capabilities look good. So, [6:35] let's get started with the build [6:36] process. [6:37] Um so, what you'll notice in a couple of [6:39] seconds here, um is that the agent is [6:41] going to or sorry, Chat GPT is going to [6:45] uh continue to live in the left-side [6:47] pane here um while it works on the right [6:49] side to uh wire up all of the, you know, [6:52] scheduling, the apps, uh the tools, and [6:55] also create a fine-tuned instruction [6:57] set. [6:58] Uh so, this is a really great no-code [7:00] way to build agents um that still, as [7:02] Christina mentioned, uh work like our [7:05] Codex uh agent. So, it can run across [7:08] long-running tasks uh behind the scenes, [7:10] access data, and take action across [7:13] systems. So, pretty shortly here, uh [7:15] we're going to see uh Chat GPT finish up [7:19] wiring up the agent, and we'll be able [7:21] to actually do a test run and show you [7:23] all of this live. [7:26] All right. So, we'll give it um a little [7:28] more time here, and I'll start to poke [7:30] around as it um finishes configuring [7:32] some of the different applications and [7:34] tools and steps. [7:39] All right. So, you can see here, um it's [7:41] finished up the instructions. Um, it's [7:44] going to, um, you know, continue to [7:46] build out the agent. At this point, I [7:48] can go in here and make modifications [7:51] directly. Um, or again, I can provide [7:53] that feedback to chat on the left side, [7:55] and it'll take care of all of that [7:57] configuration. [7:59] All right. So, almost done there. Um, [8:02] let me actually wire in some tools here. [8:05] Uh, so, for Google Calendar, and this is [8:07] helpful. I wanted to do this manually to [8:09] show you the configuration step. So, any [8:12] of the applications that you're [8:13] utilizing, um, you have the ability to, [8:16] you know, lock down what the agent [8:18] actually has access to. Uh, so, in this [8:20] case, we're looking at Google Calendar, [8:22] um, and I just need the agent to [8:24] reference my calendar. I don't need it [8:26] to take any write actions or anything [8:28] like that. Uh, so, I'm actually going to [8:30] disable, um, you know, those actions, [8:32] uh, for the agent to take those actions. [8:35] And I can know confidently that anytime [8:37] it's accessing my calendar, it's not [8:39] going to, um, you know, adjust or change [8:41] any events. Um, it's just going to go [8:44] about the workflow that I've described. [8:47] I'm also going to quickly add in Google [8:48] Drive, um, as well as Gmail. Um, [8:51] actually, it seems like [8:53] Okay, there we go. Uh, so, I'm going to [8:55] add in Gmail here. Um, you know, similar [8:58] configuration. Um, in this case, I [9:00] actually wanted to write an email to me, [9:02] uh, but maybe, you know, I can disable [9:04] some of these actions that it actually [9:05] does not need to do as a part of this [9:07] workflow. [9:09] Workspace agents is all about having the [9:11] control over what these agents are able [9:13] to do, uh, providing them a playbook, [9:16] um, and letting them execute on these [9:18] tasks on, you know, a schedule, or being [9:20] able to interface with them in tools [9:22] that your teams use today. [9:24] Let's do another configuration step [9:26] here, and I'm actually going to add in a [9:28] skill. [9:29] Um, so, we have the apps and tools that [9:32] the agent can use. Um, now, let's [9:34] actually give it some additional [9:35] direction on how to best support um, my [9:39] meeting prep creation. So, the great [9:42] thing about skills within ChatGPT and [9:44] Workspace agents is that first and [9:46] foremost, you can bring in skills from [9:48] other tools and platforms that you use [9:50] today. So, the best practices and [9:53] processes that you've codified, you can [9:55] bring them in and have your agent use [9:57] the same so that I can scale the [9:59] expertise and consistency across your [10:02] organization. [10:04] You can also choose popular skills [10:06] within your organization that already [10:07] exists or one of my favorite kind of [10:11] tools here is actually being able to [10:13] create a new skill and have ChatGPT [10:15] write it up much like it did for all of [10:17] the other agent configuration. So, let's [10:20] have it generate a skill here and add it [10:22] into this agent's playbook. [10:27] It'll kind of keep going through this. [10:29] So, we'll give it a few seconds here, [10:31] but I also have a kind of refined agent [10:34] that we can skip ahead to [10:36] and be able to jump into a preview run. [10:39] I always like to kind of observe what [10:41] ChatGPT and the agent is doing as a part [10:43] of the build process. So, we'll take a [10:45] look [10:46] at what it's doing for a little bit [10:48] longer here. [10:53] All right, so it's finding that template [10:55] within my skill or sorry, finding that [10:57] template within my Google Docs [10:59] adding it in with some formatting [11:01] instructions so that you know, and I'll [11:03] show you the final product coming up [11:06] here, but it's going to beautifully [11:07] format these meeting briefs for me with [11:10] tables, [11:11] headers, bullet points so that I can get [11:13] the information I need at a glance [11:15] especially if I'm looking at it on my [11:17] phone on the way in my morning commute. [11:21] All right, so this looks like it's [11:24] working out well. I'm actually going to [11:26] skip ahead to an agent version that's [11:28] maybe refined a bit more. You know, I've [11:31] had some time to work with it. Um and I [11:33] want to show you what it could look like [11:35] to easily run a preview test as well. [11:37] So, here um I'm going to start up a new [11:40] session. Um again, the same agent, maybe [11:43] with just a little more refinement. I've [11:45] given um ChatGPT some feedback and we've [11:48] done some preview runs. Uh so, I feel [11:50] pretty good about this agent and I'm at [11:52] that step where I can schedule and [11:53] distribute it. Uh but let's let's [11:56] actually do a quick preview test run. [11:59] So, I'm going to ask chat to do that. [12:01] Can you run a preview with my calendar [12:05] for tomorrow? [12:07] So, much like that experience of um or [12:10] the no-code experience of building these [12:12] agents, you can also work with ChatGPT [12:14] uh to actually, you know, run tests for [12:16] you. Like I said, you can give feedback [12:18] once uh these runs are executed. And you [12:21] can essentially, you know, continue to [12:23] work with it in this way. Um so, you [12:25] don't need IT and engineering support to [12:27] build out these workflows. We really [12:29] want to make sure that these subject [12:31] matter experts um are the ones that can [12:34] build out these flows and again, [12:35] distribute it to teams and share them um [12:38] as needed. So, [12:40] well, right on cue, we're going to get [12:41] this preview test uh kicked off. Um and [12:44] the other great part of these preview [12:46] runs is that you can see exactly what [12:48] the agent is doing. You can see um you [12:51] know, how it's pulling information from [12:52] other systems, um what its chain of [12:55] thought is, um and be able to follow [12:57] along. Now, this isn't going to be a [12:59] view that you baby sit or that an end [13:01] user will necessarily see every time. Uh [13:04] but it's great for those preview runs [13:05] and also show you how can how you can [13:07] look at previous [13:09] um historical runs of your agent. Um but [13:11] this is of course, especially great as [13:13] you first build out this agent and um [13:15] you do some tests before you put it into [13:18] a production setting. [13:25] All right. So, as you can see, we're [13:28] following along here um much like the [13:29] agent outline and the other ways that [13:33] chat GPT provides workspace agents [13:36] functionality. We're able to see the [13:38] high-level kind of plan for this agent [13:40] as it executes this workflow and we can [13:43] see exactly for each step you know what [13:45] it's doing, what files it's accessing, [13:47] you know how it's going about utilizing [13:49] the skills [13:50] and then we're going to get a final [13:52] result pretty shortly here that we can [13:55] review [13:56] and you know [13:57] I'll actually then show you how we can [13:59] schedule this agent, share it and [14:01] distribute it to your team for the for [14:03] them to remix and utilize for their own [14:06] workflows as well. [14:15] All right, well we'll let this run a [14:17] little longer but [14:19] in the demo industry as I like to call [14:20] it we have this concept of cooking demo [14:23] so much like on TV when someone's [14:25] putting together a meal and they pull [14:27] something out of the oven right away. [14:30] Let's actually skip ahead because I want [14:32] to be able to get to your questions and [14:33] the second demo agent as well. So if I [14:36] go into my inbox I have a previous run [14:40] that I used with this agent so you can [14:42] see here you know really great formatted [14:45] email with all the briefs that I need [14:47] and this is based off of my calendar and [14:49] all of the customer conversations that I [14:51] have for tomorrow. So it's a busy day [14:54] and I want to be able to go in and prep [14:56] for Blossom Mart, Petal Pay [14:59] as well as Nectar Works you know just at [15:03] with a lot of simplicity. So what the [15:05] agent did is it looked across my Google [15:08] Drive for all of my customer contacts, [15:10] you know maybe did some research on the [15:12] web to pull in additional information [15:15] and as you can see here it used that [15:17] skill really well to give me a highly [15:20] formatted and styled document with, you [15:22] know, the executive readout, uh customer [15:25] snapshot, meeting objective. These are [15:27] all things within my template that I [15:29] typically look for when I'm putting [15:31] together these meeting briefs. And the [15:33] great part of this is I can also share [15:34] this with my team members, so they know [15:37] exactly what the agenda will be um and [15:39] what our shared goals and outcomes uh [15:41] for the meeting will be with the [15:42] customer. So, um as you can imagine, [15:45] this is saving me, you know, hours on a [15:47] daily basis prepping for all of my [15:49] customer meetings. Um I can look at this [15:51] on the way to work or in my free time. I [15:54] don't have to do all of the manual work [15:56] um because Auto, my agent, is going to [15:58] do that for me on a scheduled basis. So, [16:01] let's go back to the agent and I'll show [16:03] the last sort of step here um of the [16:06] agent and its execution. So, what you [16:09] can do after, you know, running through [16:11] test is start to um share and distribute [16:14] this agent. So, um you can obviously add [16:17] a schedule for yourself and we would [16:18] want this to run daily for my workflow. [16:21] Um but let's say this starts to work [16:23] really well for you and everyone else [16:25] wants to have the same uh type of, you [16:27] know, efficiency and hour saving that [16:29] you're seeing with this agent. Well, [16:31] you're able to then enable uh sharing [16:34] for this agent. So, what it'll mean is [16:36] that anyone in your workspace will be [16:38] able to utilize this agent or duplicate [16:40] and remix it. Let's say, you know, [16:42] someone uses SharePoint instead of [16:44] Google Drive, they can make that change [16:46] to this agent and use ChatGPT to [16:48] reconfigure it. Um or maybe there are [16:51] additional steps you want to include [16:52] like having ChatGPT also create a slide [16:55] deck for you after uh putting together [16:58] briefs. You can um hopefully tell that [17:00] there's a lot of customize um [17:02] customization and flexibility that [17:04] Workspace agents provides to you. Um so, [17:07] hopefully you're inspired uh to build [17:09] out an agent that's similar. Um again, [17:12] you can refer back to this cookbook uh [17:14] to be able to go through uh the agents [17:16] um after this session as well. [17:20] All right. So, that was a fun one. Uh [17:22] what we did for the meeting prep agent [17:24] is we built an agent using natural [17:26] language. Uh we went into the [17:28] configuration of the app parameters and [17:30] access. Uh we enabled skills, um and [17:33] also memories. I'll include that in the [17:35] uh second agent. Um and then we ran an [17:38] agent preview, um so we can check out [17:40] the agent behavior. Um and once that [17:42] looked good, uh we decided to share it [17:44] um share the agent for team use. So, um [17:47] really helpful agent uh in my [17:49] day-to-day. Hopefully, folks um build [17:51] out your own version. But, let's go into [17:53] the second example for today. [17:56] So, this one is going to be right off [17:58] the bat uh more focused on serving the [18:01] needs of your organization or team [18:03] members. So, this is a software review [18:06] agent, and much like the previous agent, [18:08] this is actually an example that um is [18:10] modeled off a real agent that we have um [18:13] living in one of our Slack channels. I'm [18:15] sure everyone uh at your company, you [18:18] have a channel like this where [18:19] high-volume, high-sensitive, or [18:21] time-sensitive requests are made in [18:24] Slack to get a software tool by an [18:26] employee, you know, when they need it. [18:28] Um and typically, those channels, um you [18:31] know, I sometimes feel like the [18:32] experience is more like a chatbot. I [18:34] just get links. I have to do my own, you [18:36] know, discovery uh versus having an [18:39] agent take that action for you, where it [18:41] will research um the capabilities of [18:44] your requested vendor. It'll compare [18:46] options against the approved stack, you [18:48] know, reason across uh details like [18:50] utilization of existing tools, and then [18:53] it'll provide guidance to the user or [18:55] even take action. So, um you know, in [18:57] the cases that a human in the loop is [18:59] required, you can even escalate into [19:01] Jira. So, this has been an amazing agent [19:04] that internally has saved a lot of time [19:06] for our IT team, obviously, but also has [19:09] helped reduce our duplicate tool spend [19:12] and sprawl. So, really excited to go [19:14] into this example. [19:17] Um so, let me move into this experience [19:21] here. So, [19:22] um I'm going to start off from scratch [19:24] here once again just to show you that [19:26] even if we're building an agent that um [19:29] maybe takes on a little more work, uh [19:31] maybe it's a little more complex than a [19:32] personal kind of meeting prep agent that [19:35] you can still start from the same [19:37] starting point. So, here we have, you [19:40] know, a prompt and in this case I [19:42] actually have a skill that I want to [19:44] bring in. Um so, I'm going to actually [19:47] uh bring in a skill that our product or [19:50] sorry, excuse me, procurement team built [19:53] um to codify their workflow and best [19:55] practices when it comes to these uh sort [19:58] of decisions and evaluation. But, much [20:00] like the previous agent, I'm starting [20:02] off with the systems and um tools that [20:05] the agent will utilize and then I'm [20:07] giving this skill and we'll have ChatGPT [20:10] wired it all together. [20:12] Um I'll skip ahead a bit. We won't go [20:13] through the full build process cuz we [20:15] took a look at the previous one. Um but [20:18] again, uh I hope this inspires you to [20:20] start from scratch and maybe bring a [20:22] process or workflow that you're an [20:25] expert on and be able to build that uh [20:27] agentic team member that can either [20:29] assist or take over that work for you. [20:32] So, [20:34] awesome. Much like the previous agent, [20:36] we have the agent plan, you know, um [20:38] things look good. So, we'll get started [20:40] here. And maybe while this uh wires [20:43] together, I'll take the opportunity to [20:44] showcase uh the mock data set that this [20:47] agent is using. Um you know, in our [20:50] production agent it actually leverages a [20:52] software kind of management platform and [20:55] a custom MCP. Uh here I brought that [20:57] into Google Sheets just for simplicity [20:59] of the demo. Um but I pulled this up to [21:02] showcase that, you know, the agent's [21:03] actually going to look across all of the [21:06] approved software, [21:07] um you know, the utilization of uh the [21:10] tools as it is today, um and the [21:12] functionality and the capabilities of [21:14] each tool that we have, so that when a [21:16] user is requesting a tool uh for a very [21:19] specific purpose, um the agent's able to [21:22] reason and know that if it's making a [21:24] recommendation from our approved stack, [21:26] um it's actually going to meet the needs [21:28] of that user. So, that's what we have [21:29] provided to the agent. Um and what we [21:32] also have, and I'll go into, is a Slack [21:35] channel that I'll show how we can, you [21:37] know, bring that agent into Slack, uh [21:39] the interface where these requests are [21:41] being made. [21:43] So, if I go back here into Slate, um you [21:46] know, obviously, there's a lot of [21:47] configuration going on here. Um so, [21:50] maybe I'll skip ahead to [21:52] uh previously built version, but I [21:54] encourage everyone to try this out, uh [21:56] especially when you go through it for [21:57] the first time. It's one of my favorite [21:59] kind of parts of this product, uh [22:01] watching ChatGPT put everything together [22:04] and allowing you to follow along as [22:06] well. [22:08] All right. Right on cue, we have uh the [22:10] instructions. Um but yeah, I'm going to [22:12] skip ahead to that sort of refined agent [22:15] um and walk you through a couple of [22:16] setup steps before we go into Slack and [22:19] show a live test run. So, uh first and [22:22] foremost, um what we have here is uh the [22:25] ability to actually add Slack. Um so, [22:28] I'll showcase that by adding Actually, [22:31] I'll I'll pull this up and you'll be [22:33] able to see um that you can specify what [22:35] channel that an agent should operate in, [22:39] um as well as, you know, whether it [22:40] should respond when mentioned or for any [22:43] um [22:43] you know, relevant messages uh that it [22:45] can handle in the channel. [22:47] Uh you can also add additional [22:48] instructions. Um so, how to handle [22:51] requests, how to respond in thread, um [22:53] things like that. Uh so, there's a lot [22:55] of customization that you could do um [22:57] when it comes to integrating this agent [23:00] into Slack um as well. So, here, um, we [23:03] have the agent. One thing I forgot to [23:05] call on the previous agent, so I'll [23:07] bring up here, is this concept of [23:09] memory. Uh, so, memory allows your agent [23:12] to, uh, save notes, you know, contacts, [23:15] outputs, um, and other things that will [23:17] make it, uh, better at its workflow over [23:20] time. Uh, so, you can think of this as, [23:22] uh, persistent set of, uh, contacts that [23:25] the agent will keep, um, and be able to [23:27] execute on in its workflow, much like it [23:30] does with skills and tools, as well. [23:34] All right. Well, everything looks good. [23:36] Um, so, let me show you one more thing [23:38] in the platform. We'll actually go into [23:40] Slack. Um, so, on each agent screen, [23:43] you're also going to be able to see all [23:45] of its activity. So, that preview run I [23:48] showed you with the previous agent, [23:50] where you're able to see everything that [23:51] the agent did, uh, this is where, for [23:54] each agent that you create, you also [23:56] have this centralized place, uh, to view [23:59] all of the previous agent runs, um, and [24:02] you can go in here and check the agent [24:04] trace down to that level of detail. So, [24:07] um, this is especially handy, of course, [24:09] when you have an agent like this that is [24:11] autonomous and runs on its own, uh, so [24:13] that you can go back to any requests, [24:15] um, and everything is centrally audited [24:17] and, um, and captured for you. Um, and [24:20] this, of course, is also exportable via [24:22] our API. So, just want to show that [24:25] before we dive into the live example. [24:28] Uh, so, here I have my Slack channel and [24:30] some previous test runs. Um, I'm just [24:33] going to show you what it looks like to [24:34] kick off one of these, uh, cases, and [24:37] then I'll show you and jump ahead to the [24:39] agent response, uh, because it does take [24:42] a little more time, uh, than, of course, [24:43] a chatbot. Uh, the agent, like I [24:46] outlined, is actually going through, and [24:48] you can see here, the dialogue popped up [24:50] that it's working, it's received my [24:51] message, uh, but the agent is going to [24:53] go through and, um, you know, do some [24:55] research on Screen Studio, is going to [24:58] um, reason against um, you know, the [25:00] policy, the skill, um, as well as the [25:03] approved software vendor list uh, before [25:05] it gets back to the user with a [25:07] recommendation or action that it took. [25:10] So, [25:11] you know, while it's working, let's [25:12] actually go back to this one here and [25:15] I'll show you all um, you know, for the [25:17] same exact request uh, what the agent [25:19] was able to help me out with. So, Slate [25:21] responds and actually says that its [25:23] decision is that this request needs IT [25:26] review uh, versus me being able to [25:28] self-serve Screen Studio. Um, it clearly [25:31] lists out the reasoning why, the [25:33] rationale. [25:34] Um, you know, we actually have a [25:36] approved tool called Bloom um, that [25:38] would work well for my needs of needing [25:40] a high-quality uh, demo recording tool. [25:43] Uh, but it's blocked right now because [25:45] it's over utilized. There are actually [25:46] no available licenses. So, that's why [25:49] um, you know, the agent in this case has [25:52] actually chosen to escalate to IT as IT [25:55] will need to either provision additional [25:56] licenses or maybe they'll make an [25:58] exception for me um, and to use Screen [26:01] Studio in this case. Um, but you can see [26:03] that the agent is responding with the [26:05] sources checked with all of the [26:07] reasoning and um, guidance for the user. [26:09] So, I don't have to have this sort of [26:11] experience where I now have to look up [26:14] the tool, maybe provide my rationale. [26:16] The agent's done that work for me. And [26:19] as the final step here, um, you know, [26:21] instead of having requiring me to open [26:23] up my own ticket for IT review or [26:25] requiring IT to do that, it's also [26:27] created this nicely in Jira as a task [26:30] uh, for them to review and quickly [26:32] unblock me from uh, getting access to a [26:34] tool that I need. [26:36] So, um, [26:38] you know, this has been a really [26:40] powerful agent like I've noted. It's [26:42] handling all of these requests for our [26:44] team today um, and saving our IT and [26:47] procurement team a lot of time um, but [26:49] also, you know, making sure that users [26:52] when they have these time-sensitive [26:53] requests because they're typically made [26:55] last minute, uh they have the tools that [26:57] they need. [26:59] >> [gasps] [27:00] >> All right. So, with that, uh we went [27:03] into the IT agent, um how to build an [27:06] agent with existing skills, how to set [27:09] up Slack so that the agent can serve [27:11] teams within other interfaces, um how to [27:14] interact with the agent in Slack, you [27:16] know, viewing the agent run history and [27:18] traces, and then last of all, how you [27:20] can even handle escalations in Jira. So, [27:23] much like the previous agent, hopefully [27:24] this inspires folks to build out agents [27:26] that help your team members. Um but with [27:29] that, I will pass it back over to [27:32] Christina to talk through getting [27:33] started. Great. Um okay. So, now that [27:36] we've walked through some demos, I'll [27:37] cover a few few tips on how to how to [27:40] build with workspace agents. Um so, [27:42] there's really four ways to build an [27:44] agent. First, you can build one in [27:46] conversation as as we saw in the demo, [27:48] so just describe a workflow your team [27:50] already does often, and ChatGPT will [27:53] help you with the setup. Second, you can [27:55] start with one of our templates. And so, [27:57] this is a great option if you want a [27:59] faster starting point with built-in [28:00] skills and tools, and then go on and [28:02] customize from there. Um third, you can [28:04] bring in an existing workflow by [28:06] importing skills and apps from other [28:08] platforms so that you don't need to [28:10] start from scratch. And finally, if your [28:12] team is already using custom GPTs, you [28:15] can start to test those workflows in [28:16] workspace agents, um and we'll have an [28:19] automatic converter from GPTs to [28:21] workspace agents, um coming soon. [28:26] Um so, I also want to take a quick [28:29] second to talk about um GPTs and what um [28:33] this means for our GPT users. So, custom [28:35] GPTs were our first step towards [28:37] lightweight process automation, and many [28:39] teams have already found them helpful [28:41] for creating shared templates to be used [28:43] in ChatGPT. Um but when we first [28:45] launched GPTs, we didn't yet have the [28:47] right models or platform primitives to [28:48] make them truly powerful and extensible. [28:51] And so, workspace agents are the next [28:53] stage of GPTs. Shared agents that can [28:55] run multi-step workflows across tools on [28:58] schedules with approvals and [29:00] follow-through. And so, if your team [29:01] already has GPTs, I would actually start [29:04] by testing some of those workflows out [29:06] with agents. [29:08] And finally, let's talk about [29:09] permissions and admin controls. So, as [29:12] the builder, you always stay in control. [29:14] Like Hojun demoed, you decide what tools [29:16] and data your agent can use and when [29:18] approvals may be required for more [29:20] sensitive tasks. And for enterprise and [29:23] EDU plans, admins can also use [29:25] role-based access controls to define who [29:27] can use agents and which apps are [29:29] available. [29:30] And finally, the compliance API then [29:32] gives admins the ability to monitor and [29:34] manage usage over time. [29:36] So, these agents are very powerful, but [29:38] they're also built to operate within the [29:40] controls and governance that your [29:41] organization needs. [29:44] Amazing. Thank you so much, Christina [29:45] and Hojun. Um now we have some time to [29:48] answer some of the questions you've been [29:49] submitting. [29:50] Um [29:51] so, the first one that came in, quite a [29:53] few Microsoft Suite users out there. So, [29:56] maybe Hojun, can you help explain um do [29:58] agents work with Microsoft tools and a [30:01] bit about like [30:02] >> Yeah, absolutely. Yeah. So, [30:04] I did use the Google Drive mostly [30:06] because it's easier to set up for my [30:08] demos and I'm familiar with them, but [30:10] you can use Microsoft tools like [30:12] SharePoint, Outlook, and Teams. [30:15] Um and as you saw, you know, what you [30:17] can read or write or the agent can read [30:19] or write will depend on the specific app [30:21] and your Microsoft permissions as well [30:23] as your workspace admin settings, but [30:25] you'll be able to replicate what I built [30:27] with those suite of products as well. [30:29] Uh the last kind of distinction I'll [30:31] make um when it comes to the Slack [30:32] example, agents can run in ChatGPT and [30:35] Slack today. Not quite yet for Teams or [30:38] these tools, but we are actively [30:39] investing services and adding more soon. [30:43] Awesome. Thank you. [30:45] All right, the next question is about [30:47] memory. And then Christina, do you want [30:49] to talk about how agent memory is [30:50] different? Yeah, so Hojun touched a bit [30:52] on this in the demo. Um, but whenever [30:54] you build a workspace agent, you can [30:56] enable memory for that agent as well. [30:58] And so this is a persistent file system [31:00] where the agent can store files and [31:02] notes to be reused across um future [31:05] runs. Um, and you can prompt for it to [31:07] add it explicitly. It can also choose to [31:09] add it automatically. And these um this [31:12] file system is specific to every channel [31:14] that the agent is in. So if you are [31:16] using this agent in ChatGPT or you're [31:18] sharing it with someone else um in [31:20] ChatGPT, uh every user has their own [31:23] file system for their own um instance of [31:26] the agent. Um, every Slack channel also [31:28] has its own memory. So if you're in kind [31:30] of a shared Slack channel, all of the [31:32] memories within that Slack channel will [31:34] be shared across all the different [31:36] messages that are that are sent. [31:38] Very cool. Yeah. Nice. And then from now [31:40] over time is an interesting bit. [31:42] Um, cool. The other question There have [31:44] been some questions about sharing. So do [31:46] you get access automatically to all the [31:48] agents in your workspace? How do you [31:50] share them across your company? [31:52] Yeah. I can take this one. So um agents [31:55] can kept be kept private. They can be uh [31:58] specific to your workflows like I showed [32:00] with the first agent. Um, but they can [32:01] also be shared with anyone in your [32:03] organization with a link or they could [32:05] also be listed within your company's [32:07] workspace agent directory. So in that [32:10] case, um you know, these are agents that [32:12] can only be shared with members of your [32:13] ChatGPT workspace. Um, but within those [32:16] uh you're able to distribute your [32:17] agents. Uh you're able to allow others [32:20] to duplicate and remix those agents like [32:22] I was talking about with the sales [32:24] meeting prep. Um, and ultimately the [32:26] admins are able to control who's able to [32:29] build, publish, and share those agents [32:31] as well. [32:32] Um, and so for enterprise and EDU, um [32:35] that's uh within our admin settings, uh, [32:38] and our role-based access controls. [32:40] For now, we only, um, allow the owner of [32:43] the agent to actually make edits in that [32:45] agent, but very soon, [32:47] um, we'll allow multiple people to edit [32:49] the agent as well. Cool. [32:52] Awesome. [32:53] And can you create agents with different [32:55] roles? Like, any kind of best practices [32:57] for more specialized agents? [33:01] Yeah, um, I would say we just showed you [33:03] a couple, Christina alluded to the fact [33:05] that, uh, here at OpenAI, we have [33:07] agents, uh, across every single business [33:10] unit, um, and so you can have agents [33:12] with different roles. In fact, they work [33:13] best when you have a specialized role in [33:15] mind with them or for them, uh, with all [33:18] of the apps, tools, files, skills, and [33:21] instructions that they need to excel in [33:22] that role. So, [33:24] um, you can think of them as, you know, [33:26] really capable team members that you [33:27] build and you provide everything they [33:29] need to be successful, along with that [33:31] guidance. So, given that they also have [33:33] memory, like we just discussed, uh, they [33:35] can be guided or corrected over time, [33:38] uh, so they can really be improved, um, [33:40] and so, like I said, you know, get [33:42] started build these agents because it's [33:44] really easy to refine and provide [33:46] feedback, and you can continue to use [33:48] ChatGPT, [33:49] um, to, you know, change and improve [33:51] the, uh, agent behavior as well. [33:54] Awesome. Thank you. [33:55] Um, and Christina, do you have to use [33:57] the desktop app? No, um, so I mean, as [34:00] we as we demoed, you can, um, create, [34:02] run, share, or view these all directly [34:04] from the web. Um, you can use them in [34:06] ChatGPT, you can use them in Slack, um, [34:09] and they run in the cloud, so they keep [34:10] working even when you're away. Um, you [34:13] can also use them from the desktop app, [34:15] you can use them from mobile as well, so [34:17] really kind of across our different [34:18] surfaces. Yeah. Yeah, I'll just add to [34:21] that that both the agents that we [34:23] demonstrated today, um, aren't actually [34:25] running on the app or locally, they're [34:26] both in the cloud. Uh, one's running on [34:29] the schedule, um, and giving me notes [34:31] for my commute, and then the second one [34:33] obviously is living in Slack. [34:35] Awesome. [34:37] Cool. And Hojun, could you answer [34:38] questions about pricing? Yeah, [34:40] absolutely. So, um, pricing uh, right [34:43] now Workspace agents are in research [34:45] preview, and they're uh, free uh, until [34:48] May 6th. So, you know, I'm sound like a [34:50] broken record, but uh, when I say to [34:52] when I'm encouraging all to try it, you [34:54] know, really put it through its paces [34:55] and try it out during that period. Um, [34:58] after that, usage will move to [34:59] credit-based pricing. Um, so the number [35:02] of credits that an agent run consumes, [35:04] um, will obviously depend on the [35:06] complexity of the agent and the tasks [35:07] that it performs. Um, but you can think [35:09] of it similar to how we, um, or more [35:12] complex Codex tasks like you uh, that [35:15] use more credits. Um, because again, [35:17] these are agents that have the same [35:18] harness and the ability to perform uh, [35:21] long-running complex tasks. Uh, we'll be [35:24] sharing more specific guidance uh, [35:26] through admin and billing channels [35:27] regarding pricing as we get closer to [35:29] uh, on that May 6th date. [35:31] Awesome. And then we actually had a [35:33] couple more questions come in. Um, so [35:35] since you mentioned Codex, Christina, [35:37] maybe you could talk a little bit about [35:39] what makes Workspace agents different [35:41] than working with Codex? Yeah, [35:43] definitely. So, I think, um, first of [35:45] all, they're they're meant for teams, [35:46] and so they can be shared with, um, [35:48] people in different functions at your [35:50] company. Um, they also run in the cloud, [35:53] and so they run even when your when your [35:54] computer is closed. Um, and then, yeah, [35:57] again again, they can be kind of [35:58] deployed in these different environments [36:00] as well. So, used directly from Slack [36:01] with, um, many other triggers coming [36:03] soon. [36:05] Awesome. And then quite a few questions [36:07] around skills. Um, can you help folks [36:09] understand, you know, what's the [36:10] difference between skills and agents? [36:12] Can agents run multiple skills? And how [36:14] do you kind of determine, um, when you [36:16] should add a skill versus defining the [36:19] agent instructions? [36:20] >> Yeah, absolutely. So, I would think [36:22] about skills as being sort of, you know, [36:24] playbook, best practices, and processes [36:27] that you have that you can codify or [36:29] maybe you're already using in different [36:31] platforms and tools. [36:33] And the agent is the worker that [36:35] actually leverages that skill and other, [36:37] you know, instructions that you have to [36:39] do the real work. So, an agent can have [36:42] multiple skills and be able to utilize [36:45] them for the specific, you know, task [36:47] that that it's working on. [36:49] So, you can think of it in that way. [36:50] Once again, skills are going to be like [36:52] your best practices. Sometimes there are [36:54] scripts, there are other [36:57] kind of direction and guidance, and then [36:59] the agent will actually take on the [37:00] scale, will use also the apps and tools, [37:03] and execute on that work [37:06] as a whole. [37:07] Awesome. [37:08] Thank you so much for taking these [37:10] questions. [37:11] Yeah, I think we're going to wrap up and [37:13] thank you everyone so much for joining [37:14] today. So, to wrap up here, just a few [37:17] helpful resources that we will share [37:19] after this session. [37:21] So, like Hojun mentioned, the cookbook [37:23] for building the sales meeting prep, as [37:25] well as some other resources with [37:27] additional demos. [37:29] And then also want to do a quick plug on [37:31] upcoming build hours. We have two [37:33] upcoming sessions that are focused on [37:35] our API. So, you can just follow the [37:37] link here to watch [37:38] to sign up for those sessions and also [37:40] watch past build hours. [37:43] Thank you so much for attending, and [37:46] thank you so much to Christian and Hojun [37:47] again. [37:48] Yeah. Thanks all. Thank you. Thank you.