OpenAI Deploys Autonomous Cloud Agents to Run Enterprise Workflows

OpenAI

Delegating routine research and communication triage to persistent cloud agents allows workers to reallocate scarce attention toward high-leverage strategic initiatives.

AI-assisted summary of How OpenAI Puts ChatGPT to Work | DevDay 2026 — original source. Check the original for full context.

Source published: 2026-10-09

Key ideas

OpenAI highlighted ChatGPT Work and "Dots"—persistent, autonomous cloud agents running on dedicated virtual machines with their own browsers to execute recurring operational workflows independently. (14:21)
OpenAI introduced ChatGPT Spaces, where teams and AI agents co-edit live collaborative pages governed by repository-style "page instructions" similar to an agents.md file. (12:04)
A two-person competitive research team at OpenAI codified its investigative methodology into a dedicated agent plugin and staging portal, automating deep-dive analyses across sales transcripts, analyst calls, and podcasts. (6:38)
OpenAI's Chief People Officer encoded internal organizational research into an interactive, voice-driven leadership coach powered by GPT Live for real-time strategic guidance. (10:29)

Detailed notes

## Background Published on YouTube by OpenAI on October 7, 2026, this DevDay 2026 session explores how internal teams at OpenAI put ChatGPT to work across everyday operations. According to the description, the presentation demonstrates how teams manage high-profile launches, conduct competitive research, and convert domain expertise into reusable tools that viewers can apply to their own workflows.

## TL;DR Presenters Nick and Danielle demonstrate how OpenAI employees overcome daily operational chaos by deploying ChatGPT as an agentic collaborator rather than a standard prompt-and-response tool. Through internal case studies—including Dani automating product launch social triage, Adrian scaling competitive intelligence, and Casey turning leadership research into a voice-powered coach—the session highlights new collaboration patterns using ChatGPT Spaces and autonomous cloud agents called 'Dots.' These systems manage background tasks like web scraping, transcript analysis, and notification filtering, allowing small teams to achieve outsized leverage. Ultimately, the speakers urge knowledge workers to delegate routine operational burdens so they can reinvest their reclaimed attention into creative and strategic pursuits.

## Atmosphere & style The presentation balances the polished energy of a flagship tech conference with the relatable, self-effacing humor of navigating daily workplace burnout. Nick and Danielle share an effortless, upbeat chemistry, warming up the room with nods to morning chaos before diving into real internal workflows. The narrative arc moves briskly from everyday overwhelm to empowered autonomy, making cutting-edge agentic workflows feel approachable, tactical, and quietly transformative.

## Bulletin - [14:21](https://www.youtube.com/watch?v=SXCfs7mitPo&t=861s) OpenAI highlighted ChatGPT Work and "Dots"—persistent, autonomous cloud agents running on dedicated virtual machines with their own browsers to execute recurring operational workflows independently. - [12:04](https://www.youtube.com/watch?v=SXCfs7mitPo&t=724s) OpenAI introduced ChatGPT Spaces, where teams and AI agents co-edit live collaborative pages governed by repository-style "page instructions" similar to an agents.md file. - [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) A two-person competitive research team at OpenAI codified its investigative methodology into a dedicated agent plugin and staging portal, automating deep-dive analyses across sales transcripts, analyst calls, and podcasts. - [10:29](https://www.youtube.com/watch?v=SXCfs7mitPo&t=629s) OpenAI's Chief People Officer encoded internal organizational research into an interactive, voice-driven leadership coach powered by GPT Live for real-time strategic guidance.

## Key takeaways - [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) Morning operational triage across inboxes and messaging apps drains cognitive energy, preventing knowledge workers from focusing on high-impact strategic tasks. - [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) ChatGPT Work shifts the paradigm from prompt-and-response transactional assistance to an agentic teammate that pursues high-level objectives with broad contextual awareness. - [03:51](https://www.youtube.com/watch?v=SXCfs7mitPo&t=231s) Delegating real-time feed monitoring and content logging to agents enables proactive, self-updating executive reporting dashboards that eliminate ad-hoc data requests. - [03:51](https://www.youtube.com/watch?v=SXCfs7mitPo&t=231s) AI agents acting as silent observers in team chat channels can filter high-volume noise and escalate only critical, actionable items directly to human owners. - [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) Codifying analytical methodologies into custom plugins allows tiny teams to handle enterprise-scale inquiry volume across product, sales, and strategy organizations. - [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) Autonomous agents can ingest and synthesize unstructured inputs—such as 90 days of sales transcripts and audio podcasts—to rigorously stress-test and invalidate internal hypotheses. - [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) Pairing automated background investigation with human-in-the-loop review staging ensures high research rigor before democratizing insights across self-serve company portals. - [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) Automating routine research tasks liberates domain experts to create sophisticated downstream tools, such as interactive benchmark and cost visualization portals. - [10:29](https://www.youtube.com/watch?v=SXCfs7mitPo&t=629s) Static internal documentation can be converted into interactive, voice-enabled coaching systems powered by GPT Live to scale specialized organizational expertise. - [11:27](https://www.youtube.com/watch?v=SXCfs7mitPo&t=687s) Adding explicit instructions directly to shared knowledge pages ensures that AI agents editing documents adhere strictly to required formatting, operational norms, and tone. - [12:04](https://www.youtube.com/watch?v=SXCfs7mitPo&t=724s) ChatGPT Spaces allow team members to run isolated private chat investigations that push verified findings back into a unified, collaboratively edited document. - [14:21](https://www.youtube.com/watch?v=SXCfs7mitPo&t=861s) Dedicated cloud agents ("Dots") maintain persistent state across modalities, allowing seamless interaction via web browsers, Slack, SMS, and phone calls. - [14:21](https://www.youtube.com/watch?v=SXCfs7mitPo&t=861s) Autonomous agents do not require exhaustive manual configuration on day one, progressively adapting to individual user habits and operational preferences over time. - [16:45](https://www.youtube.com/watch?v=SXCfs7mitPo&t=1005s) Treating personal attention as finite capital and deliberately automating familiar responsibilities creates the necessary bandwidth to pivot toward higher-value initiatives.

## Detailed walkthrough ### [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) Morning Chaos and the Promise of ChatGPT Work

Nick and Danielle from OpenAI's developer experience team open their session with an interactive show of hands. Nick first asks the audience to raise their hands if, upon starting their day, their inbox is completely clean, their mind is crystal clear, and they know exactly what needs to get done. Only one or two hands go up in the entire room. He then asks how many people feel like opening their laptop is "like a clown has jumped out of your laptop" or "like a jack in the box" of utter chaos—prompting a flood of raised hands across the audience.

Nick walks through that familiar spiral: checking emails, catching up on endless Slack threads, scanning dashboards, and trying to prioritize a to-do list where "everything" feels urgent. By the time that initial triage wraps up, half the morning has vanished before actual, meaningful work has even begun. While everyone recognizes that ChatGPT has the potential to transform this dynamic, the core question is *how*.

To answer this, Nick and Danielle interviewed internal "power users"—prolific outliers within their respective organizations at OpenAI who took on workloads that seemed humanly impossible just months earlier. They introduce three internal champions: - **Dani**, who leads social for OpenAI developers. - **Adrian**, who leads competitive research. - **Casey**, Chief People Officer at OpenAI.

While people already rely on ChatGPT as a thought partner and Codex as a coding agent, Nick frames **ChatGPT Work** (or simply "Chat") as an agentic teammate designed to get things done. Rather than merely answering prompts, it accepts high-level objectives paired with context—such as ingesting a full DevDay agenda and generating an interactive recommendation site for attendees. By handing off routine operational responsibilities and eliminating bottlenecks, professionals can finally reclaim time for projects they have put off waiting for "when things slow down."

---

[03:51](https://www.youtube.com/watch?v=SXCfs7mitPo&t=231s) Dani: Automating Launch Triage, Social Feeds, and Executive Updates

Shepherding product launches through the developer ecosystem requires constant monitoring of community sentiment and rapid stakeholder reporting. For Dani, those stakeholders include executive leadership like CEO Sam Altman asking how a high-profile release (like the hypothetical launch of "GPT-6 Astra") is performing. Previously, Dani faced hours of manual scrolling across social feeds and internal messaging apps just to separate actionable signal from noise.

To streamline this, Dani delegated ongoing monitoring to her custom AI agent ("dot"). Her agent reads the social feeds in real time, logs posts along with live engagement metrics into a central content tracking spreadsheet, and populates an internal live reporting site. When executives ask for status updates, Dani shares an always up-to-date dashboard rather than spending her day pulling ad-hoc numbers.

```mermaid flowchart TD accTitle: Dani's Social Monitoring and Reporting Workflow accDescr: Incoming social posts and Slack channels are monitored by an agent, feeding a content metrics sheet, an executive dashboard, and urgent alerts.

subgraph Ingestion["Input Feeds"] A["Social Media Feeds"] B["Launch Slack Channels"] end

subgraph AgentAction["Dot Agent Processing"] C["Agent (Dot) Monitors & Filters Signal"] end

subgraph Outputs["Outputs & Handoffs"] D["Live ChatGPT Launch Site\n(Executive Visibility)"] E["Content Sheet\n(Metrics & Performance Tracking)"] F["Urgent Slack Pings\n(Direct to Dani)"] end

A --> C B --> C C --> D C --> E C --> F ```

The dot also monitors Dani's flood of launch-related Slack channels, pinging her only when an urgent issue arises or when her direct input is needed. With data aggregation, content analytics, and channel triage offloaded, Dani can spend her energy on strategic decisions: determining how OpenAI should respond, translating developer feedback back to product engineering and developer experience teams, and directly spotlighting the work of the developer community.

- **Proactive Executive Reporting**: Generates a self-updating dashboard so updates are ready before leadership asks. - **Engagement Analytics**: Automatically archives posts and tracks engagement metrics to analyze what content performs best. - **Notification Triage**: Filters dozens of busy launch Slack channels and alerts Dani only on urgent action items.

---

[06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) Adrian: Scaling Competitive Research from a Two-Person Team

Adrian leads OpenAI’s competitive research team, operating with just two people while servicing inquiries from sales, product, go-to-market (GTM), and strategy across the entire company. A single inquiry about a competitor’s product shift or revenue spike can demand days of deep research, meaning a two-person team historically had to leave countless inquiries unanswered.

To overcome this constraint, Adrian codified his entire investigative methodology into a dedicated plugin. The plugin replicates how he approaches research: gathering disparate data points, cross-referencing sources, spotting patterns, and actively stress-testing hypotheses so conclusions survive intense internal scrutiny.

In one instance, Adrian investigated a competitor's surging revenue to see if they were landing massive enterprise accounts. His dot combed through analyst calls, industry newsletters, and tech podcasts, proving the large-customer thesis was unsupported. In another investigation, Adrian directed the agent to analyze 90 days of sales call transcripts, uncovering real drivers behind lost deals that account executives had failed to capture in standard CRM summaries.

```mermaid flowchart LR accTitle: Adrian's Competitive Research Pipeline accDescr: Questions from Slack or SMS are processed by a research plugin, staged for human review, and published to Compete Corner.

subgraph Inputs["Inbound Questions"] Q1["Slack Inquiries"] Q2["SMS Forward\n(During Commute)"] end

subgraph Analysis["Agent Investigation"] P["Research Plugin\n(Cross-checks Transcripts, Podcasts, Calls)"] S["Staging Space\n(Findings & Source Citations)"] end

subgraph Review["Human-in-the-Loop & Publishing"] H["Adrian Review\n(Adds Context / Approves)"] C["Compete Corner Site\n(Company-Wide Self-Service)"] end

Q1 --> P Q2 --> P P --> S S --> H H -->|"Approved"| C ```

To democratize these insights, Adrian created **Compete Corner**, a self-serve portal organizing research around recurring company questions. His dot monitors Slack for competitive inquiries, conducts research via his plugin, and prepares findings in a staging space with full source attribution. Adrian or his teammate reviews the conclusions, layers in extra context, and approves publication to Compete Corner. Adrian can even forward a competitor announcement via SMS during his morning commute; by the time he arrives at his desk, the complete analysis is staged and waiting for review.

With routine research automated, Adrian had time to build **Model Galaxy**, an interactive visualization tool powered by GPT-6 Astra that tracks performance, latency, and cost across frontier model benchmarks.

---

[10:29](https://www.youtube.com/watch?v=SXCfs7mitPo&t=629s) Casey: Encoding Organizational Expertise into a Live Leadership Coach

Casey, OpenAI's Chief People Officer, focuses on how AI alters organizational structures and team dynamics. Studying teams both inside and outside the AI industry, his group generates an extensive archive of research papers, meeting notes, and case studies into a single unified knowledge base.

Rather than letting that research sit as static documentation, Casey transformed it into an interactive coaching site for leadership powered by GPT Live. Leaders can have real-time voice conversations with the coach, discussing workflow bottlenecks and discovering how their teams can restructure work or delegate tasks to AI agents. Casey's team treats this system as an evolving, shared resource, designed to keep internal coaching aligned with emerging operational insights.

[11:27](https://www.youtube.com/watch?v=SXCfs7mitPo&t=687s) Scaling Expertise with Pages and Shared Instructions

The speakers explain how teams consolidate scattered notes and individual findings into a single shared resource. Rather than keeping insights siloed across personal threads, team members pool their meeting notes and research onto a unified page. By adding explicit instructions directly to the page, they define how new research should be organized, ensuring that when an AI agent works within that document, it automatically adheres to a consistent structure and style.

This centralized repository acts as a real-time knowledge base that powers downstream applications, such as internal coaching tools. As team members update the page with emerging insights, the AI coach seamlessly absorbs that material and reflects it in its user suggestions. The presenters pose an open-ended question to the audience: "What are you an expert on? And how could you make that accessible to more people?"

[12:04](https://www.youtube.com/watch?v=SXCfs7mitPo&t=724s) Introducing ChatGPT Spaces: Real-Time Team Collaboration

Turning to concrete workflows, the presenters highlight the newly launched ChatGPT Spaces, showing how internal OpenAI champions structure their day-to-day operations. Taking the earlier example of Adrian's "Compete Corner," they reveal the underlying space that hydrates the entire competitive intelligence site. Rather than generating reports in isolation, ChatGPT dumps its findings directly into Adrian's collaborative space, where multiple team members can review, edit, and comment simultaneously on the exact same page.

```mermaid flowchart TD accTitle: Knowledge Flow in ChatGPT Spaces accDescr: Diagram illustrating how private chats, shared page instructions, and team edits converge into shared pages and downstream tools. subgraph CollaborativeSpace["ChatGPT Space"] direction TB Agent["AI Research Agent"] Instructions["Page Instructions (agents.md style)"] Page["Shared Page (Live Collaborative Document)"] UserA["Adrian (Private Chat)"] UserB["Teammate (Private Chat)"] end

Output["Downstream Tools (e.g., Compete Corner Site)"]

Inside this environment, team members enjoy both shared collaboration and private exploration: - Teammates question claims, leave inline comments, and drop source links while working simultaneously on the page. - Each person can spin up their own private chat thread within the space to run side investigations without cluttering the common view, yet any resulting edits are written directly back to the shared page. - Spaces incorporate "page instructions"—comparable to an `agents.md` file in a software repository—that enforce operational norms, working practices, and voice guidelines for any agent editing that page. - Beyond text pages and file storage, spaces will soon support generating slides, spreadsheets, and additional workspace artifacts directly within ChatGPT.

To get started, the presenters suggest taking an existing lengthy ChatGPT thread and asking the model to summarize or convert it into a page. From there, users can pair directly with their agent to refine the text before inviting colleagues to co-edit in real time.

[14:21](https://www.youtube.com/watch?v=SXCfs7mitPo&t=861s) Dots: Autonomous Cloud Agents and Everyday Hand-offs

The session transitions from collaborative documents to autonomous execution with the launch of "Dots"—personal AI sidekicks designed to track evolving workflows in the background. While spaces organize static context, a Dot acts as an always-on, proactive operator running on its own dedicated virtual machine in the cloud, complete with its own browser.

```mermaid flowchart LR accTitle: Dot Architecture and Touchpoints accDescr: Shows the Dot cloud core connecting to input channels, web tools, and alert destinations. subgraph DotCloud["Dot (Cloud Container)"] Core["Persistent Agent Context"] Browser["Cloud Browser / VM"] end

subgraph Inputs["Interactions"] Phone["Phone Call / SMS"] Desk["Laptop Files"] SlackIn["Slack Messages"] end

subgraph Actions["Autonomous Tasks"] Social["Scrape Feeds / Web"] Sheet["Update Content Sheets"] SlackAlert["Ping User on Slack"] end

Inputs --> Core Core --> Browser Browser --> Social Browser --> Sheet Core --> SlackAlert ```

The speakers illustrate how Dani handed off launch monitoring to her Dot: - **Autonomous tracking:** Dani's Dot independently navigates social media feeds via its cloud browser, gathers posts, and updates her content spreadsheet without human intervention. - **Flexible participation:** A Dot can join Slack channels as a silent observer ("a fly on the wall") or as an active participant, pinging the team only when urgent items require human review. - **Persistent presence:** Because the Dot lives in the cloud, it maintains continuous context across modalities. Users can speak with it over phone calls, message it during commutes, interact via Slack, or give it access to local laptop files.

The speakers walk through setting up a Dot: giving it a name (one speaker named hers "Wyatt" after a childhood dog, while another chose "Ginger" after the Spice Girl), customizing its visual avatar with accessories (such as a frog face with a monocle), assigning its core responsibilities ("tell it what you want, what you really, really want"), and linking preferred notification channels. They note that user preferences do not need to be configured exhaustively on day one, as the Dot progressively adapts to working styles over time. Attendees are also directed to visit the physical "Dot shop" downstairs at DevDay to customize their own sidekick.

[16:45](https://www.youtube.com/watch?v=SXCfs7mitPo&t=1005s) Repurposing Attention and Personal Automations

The session wraps up with an audience exercise and a personal reflection on what AI delegation enables. Displaying an on-screen QR code, the speakers invite the audience to pull out their smartphones to open a guided ChatGPT onboarding experience. The tool prompts users about their day-to-day duties, attempts to infer their workflow, and recommends their first task to hand off using Spaces and Dots.

Reflecting on the broader impact of AI, one presenter recounts his career path at OpenAI: - He originally started on the technical onboarding team, tasked with getting incoming software engineers up to speed on OpenAI's infrastructure and architecture. - By deliberately automating large portions of that curriculum and onboarding support using ChatGPT, he created the room needed to transition into developer experience. - Stepping away from tasks he was already comfortable doing felt "a little bit strange," but ultimately enabled him to pivot toward work he found truly exciting.

Closing the presentation, he frames time and mental energy as modern capital: "Attention is currency and you get to decide where to spend it. So, let me ask you: What will you hand off? And what could you finally take on?" Attendees are invited to connect with the presenters and Dani afterward for follow-up questions.

## Notable quotes > [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) like a clown has jumped out of your laptop — Nick

> [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) like a jack in the box — Nick

> [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) when things slow down — Nick

> [11:27](https://www.youtube.com/watch?v=SXCfs7mitPo&t=687s) What are you an expert on? And how could you make that accessible to more people?

> [14:21](https://www.youtube.com/watch?v=SXCfs7mitPo&t=861s) a fly on the wall

> [14:21](https://www.youtube.com/watch?v=SXCfs7mitPo&t=861s) tell it what you want, what you really, really want

> [16:45](https://www.youtube.com/watch?v=SXCfs7mitPo&t=1005s) a little bit strange — Nick

> [16:45](https://www.youtube.com/watch?v=SXCfs7mitPo&t=1005s) Attention is currency and you get to decide where to spend it. So, let me ask you: What will you hand off? And what could you finally take on? — Nick

## Mentioned OpenAI, DevDay 2026, Nick, Danielle, Dani, Adrian, Casey, Sam Altman, ChatGPT, ChatGPT Work, Codex, GPT-6 Astra, GPT Live, Compete Corner, Model Galaxy, ChatGPT Spaces, Dots, Slack

## Video description See how OpenAI’s teams use ChatGPT to manage launches, research competitors, and turn expertise into tools. Take practical ideas back to your own work.

[Source](https://www.youtube.com/watch?v=SXCfs7mitPo)

中文简报

OpenAI 部署自主云端智能体运行企业工作流将日常调研与沟通分类整理委托给常驻云端智能体，使员工能够将稀缺的注意力重新分配给高杠杆战略举措。演讲者 Nick 和 Danielle 展示了 OpenAI 员工如何通过将 ChatGPT 部署为智能体协作者、而非传统的单次提示问答工具，来化解日常运营混乱。通过内部案例研究——包括 Dani 实现产品发布社交媒体信息分流自动化、Adrian 扩展竞品情报工作，以及 Casey 将领导力研究转化为语音驱动的教练——该演讲重点介绍了结合 ChatGPT Spaces 与被称为“Dots”的自主云端智能体的新协作模式。这些系统负责处理网页抓取、文本转录分析和通知筛选等后台任务，让小团队能够发挥巨大的杠杆效益。最后，演讲者呼吁知识工作者下放日常运营负担，从而将释放出的精力重新投入到创造性和战略性事务中。 OpenAI 重点介绍了 ChatGPT Work 和“Dots”——这些常驻、自主的云端智能体运行在专用虚拟机上，并拥有自己的浏览器，能够独立执行周期性运营工作流。 (14:21) OpenAI 推出了 ChatGPT Spaces，团队和 AI 智能体可在其中协同编辑实时协作页面，并由类似于 agents.md 文件的代码仓库式“页面指令”进行管控。 (12:04) OpenAI 的一个两人竞品调研团队将其研究方法固化为专用智能体插件和暂存门户，实现了对销售记录、分析师电话会议和播客内容的自动化深度分析。 (6:38) OpenAI 首席人才官将内部组织研究转化为由 GPT Live 驱动的交互式语音领导力教练，以提供实时战略指导。 (10:29)

Reading summary

At DevDay 2026, OpenAI's developer experience team presented ChatGPT Work as an agentic operational partner designed to eliminate daily triage bottlenecks by delegating complex workflows to collaborative environments and autonomous cloud agents. Automating Social Triage and Executive Reporting (3:51) Managing developer product rollouts historically required hours of manual triage across internal channels and public feeds to distinguish meaningful signal from noise. To support releases like the hypothetical GPT-6 Astra launch, Dani, OpenAI's social lead for developers, transferred routine tracking to an autonomous agent named dot. The agent continuously monitors social platforms in real time, aggregates engagement metrics into a central tracking spreadsheet, and populates a live reporting dashboard accessible directly by executive leadership like Sam Altman. Scaling Research Pipelines via Codified Methodologies (6:38) Adrian's competitive intelligence group supported company-wide requests from sales, product, and strategy with only two team members, creating significant operational bottlenecks. To scale capacity, Adrian encoded his investigative approach into a dedicated research plugin that synthesizes disparate data sources—including analyst calls, podcasts, and 90 days of sales call transcripts—to test business hypotheses, such as debunking whether a competitor's revenue surge stemmed from enterprise wins. Collaborative Knowledge Synthesis in ChatGPT Spaces (12:04) To prevent institutional expertise from remaining trapped in isolated chats, OpenAI introduced ChatGPT Spaces, which allow teams to organize research onto shared, live pages. Teams incorporate standardized page instructions—analogous to an agents.md file in software development—to enforce tone, structure, and research guidelines for editing agents. Inside these spaces, colleagues can leave comments and run parallel private chat threads that update the common document directly, hydrating self-service resources like Adrian's Compete Corner and Casey's GPT Live executive coaching portal. Autonomous Cloud Execution with Persistent Dots (14:21) Beyond shared workspaces, OpenAI showcased 'Dots'—customizable, always-on AI assistants that operate on dedicated cloud virtual machines equipped with their own browsers. Operating across modalities including phone calls, SMS, laptop files, and Slack, a Dot can act as a silent observer or an active participant that executes tasks like web scraping and alerts users only when urgent human intervention is required. The presenters frame attention as personal currency, illustrating how deliberately automating established onboarding and support tasks allowed team members to pivot toward strategic work. However, adopting these autonomous agents requires teams to codify operational norms into structured instructions and overcome the discomfort of handing off familiar responsibilities.

Detailed notes — zh

## Background 本视频由 OpenAI 官方频道于 2026 年 10 月 9 日在 YouTube 平台发布，是 DevDay 2026 的专题演讲，总时长 19 分 41 秒。视频简介指出，内容主要展示 OpenAI 内部团队如何利用 ChatGPT 管理产品发布、调研竞争对手以及将专业知识转化为实用工具，旨在为观众提供可带回自身工作场景的实践思路。 ## TL;DR 来自 OpenAI 开发者体验团队的 Nick 和 Danielle 分享了内部员工如何将 ChatGPT 从单纯的“头脑风暴伙伴”升级为“具备代理能力的团队协作伙伴（agentic teammate）”。演讲通过三个核心案例展开：社交媒体负责人 Dani 利用智能体实时监控产品发布舆情并自动更新高管看板；竞品研究负责人 Adrian 将调研方法论沉淀为插件并构建内部情报站，以两人团队撬动全公司的研究需求；首席人才官 Casey 则整合零散调研打造了基于语音的领导力 AI 教练平台。此外，演讲还重点演示了支持多人与多智能体协同的“ChatGPT Space”以及全天候云端值守的智能体副手“Dots”，鼓励知识工作者将繁琐的例行事务交接给 AI，从而将宝贵的注意力投入到更具创造力和高价值的事务中。 ## Atmosphere & style 整场演讲氛围轻松幽默且极具启发性，两位讲者 Nick 和 Danielle 配合默契，通过关于“清晨收件箱混乱”的现场举手互动迅速拉近了与台下开发者的距离，引发了全场会心的苦笑与共鸣。叙事节奏从日常工作痛点切入，层层递进到具体的内部高管案例与硬核新功能演示，其间穿插着为智能体起宠物狗名字或“辣妹组合”名字等生活化趣事，显得亲切自然。演讲尾声通过个人职业转型的真实经历与“注意力即货币”的金句升华主题，给人留下一股摆脱繁琐日常、重获工作掌控感与创造力的温暖鼓舞。 ## Bulletin - [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) 在企业工作场景中，ChatGPT 的核心定位已从单纯的“头脑风暴伙伴”演进为“具备代理能力的团队协作伙伴（agentic teammate）”，能够长期接管持续性的例行任务并解放员工精力。 - [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) 通过将研究方法论、知识库与日常监控交由 AI 处理，OpenAI 内部的社交运营、双人竞品研究及人才战略团队得以摆脱繁琐的信息搬运，实现高杠杆产出与专业知识的工具化分发。 - [12:14](https://www.youtube.com/watch?v=SXCfs7mitPo&t=734s) OpenAI 推出全新协作空间 ChatGPT Space，支持多成员与多 Agent 在同一页面协同作业，在保留个人私密对话追问的同时，可通过页面指令（Page Instructions）统一团队沉淀标准。 - [14:33](https://www.youtube.com/watch?v=SXCfs7mitPo&t=873s) 全天候智能体副手“Dots”正式亮相，在云端拥有独立电脑与浏览器环境，可通过 Slack、短信与语音通话跨端无缝交互，实现全天候自主监控与按需提醒。 ## Key takeaways - [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) 知识工作者清晨常陷入邮件、Slack 消息与业务仪表盘的混乱中，而将 ChatGPT 作为“具备代理能力的团队协作伙伴”，只需赋予明确的目标意图与上下文即可交接持续性例行工作。 - [03:51](https://www.youtube.com/watch?v=SXCfs7mitPo&t=231s) 开发者社交负责人 Dani 借助专属智能体（Dot）实时读取社交媒体流与 Slack 频道动态，自动填充发布看板与内容复盘表格，使自身角色从被动的“舆情搬运工”转变为深度策略制定者。 - [03:51](https://www.youtube.com/watch?v=SXCfs7mitPo&t=231s) 面对管理层随时可能发起的业务追问，提前通过智能体持续汇总实时指标与开发者情绪反馈，能让关键洞察时刻处于“准备就绪”状态。 - [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) 仅有两人的竞品研究团队将调查、模式识别与交叉验证的方法论沉淀为定制插件，横向穿透财报电话会、分析师通讯和播客，用客观事实推翻了业内对某竞品收入激增的表面假设。 - [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) 智能体通过自动复盘过去 90 天内的所有销售沟通录音文本，成功挖掘出销售人员在 CRM 摘要中未曾记录的真实丢单归因，拼凑出完整的客户流失图谱。 - [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) 借助内部竞品知识门户“Compete Corner”与基于 GPT-6 Astra 开发的“Model Galaxy”交互式可视化系统，两人小团队实现了全公司自助查验竞对论据与前沿模型多维指标对比。 - [10:29](https://www.youtube.com/watch?v=SXCfs7mitPo&t=629s) 首席人才官 Casey 团队将海量会议纪要、学术论文与行业报告整合为共享知识底座，并依托 GPT Live 打造语音交互的领导力 AI 教练平台，协助管理者梳理业务瓶颈与 AI 任务分配。 - [11:27](https://www.youtube.com/watch?v=SXCfs7mitPo&t=687s) 通过在共享页面设定页面指令（Page Instructions）规范结构与风格，团队能将分散的个人探索凝聚为统一的组织级资产，并直接作为底层上下文反哺给 AI 教练站点实现动态更新。 - [12:14](https://www.youtube.com/watch?v=SXCfs7mitPo&t=734s) ChatGPT Space 实现了“私密对话与公共协同解耦”，团队成员可在各自私密的对话线程中向 AI 追问细节，再将提炼后的结论与佐证资料实时同步至公共页面，并将在未来原生支持幻灯片与电子表格。 - [14:33](https://www.youtube.com/watch?v=SXCfs7mitPo&t=873s) 智能体副手 Dot 具备独立的云端虚拟电脑和内置浏览器，不仅能自主浏览网页和更新排期表格，还可入驻 Slack 频道充当后台观察者或讨论参与者，仅在检测到紧急事务时呼叫人类介入。 - [14:33](https://www.youtube.com/watch?v=SXCfs7mitPo&t=873s) Dot 拥有全天候在线能力与跨终端的一致性记忆，无论用户通过手机短信、语音通话、Slack 还是桌面端与其交互，背后的智能体始终保持上下文连贯，并能在日常协作中持续学习用户习惯。 - [16:48](https://www.youtube.com/watch?v=SXCfs7mitPo&t=1008s) “注意力就是货币”，将繁复的基础工作（如技术入职培训、枯燥资料收集与日常盯盘）自动化并交接给 AI，能让人类员工夺回注意力的支配权，转身投身于更具创造力与核心价值的事业。 ## Detailed walkthrough ### [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) 晨间混乱与 ChatGPT 的工作新角色演讲在一场针对现场开发者的互动调查中拉开序幕。来自 OpenAI 开发者体验团队的 Nick 与 Danielle 走上讲台，向台下抛出第一个问题：“每天清晨开启工作时，有谁的收件箱是完全清空的，内心无比笃定，清晰地知道今天每一步该做什么？”台下寥寥无几，仅有一两个人举手。紧接着，Nick 问出了相反的场景：“又有谁在掀开笔记本电脑屏幕的瞬间，感觉就像有一个小丑突然从盒子里蹦了出来（jack-in-the-box），满眼都是无法招架的混乱？”这一次，台下几乎所有人都在苦笑中举起了手。这种工作现状每个知识工作者都不陌生：打开电脑，先浏览堆积如山的邮件，爬完 Slack 里的未读消息，再扫一眼各大业务仪表盘；当你终于理清头绪准备推进核心任务时，却发现“一切都迫在眉睫”。转眼间半个上午已经耗尽，而真正的本质工作甚至还没正式开始。Nick 和 Danielle 指出，ChatGPT 正在彻底重构这一局面。为此，他们深入访谈了 OpenAI 内部最顶尖的三位“ChatGPT 重度超级用户”——负责开发者社交媒体运营的 Dani、主导竞品研究的 Adrian，以及 OpenAI 首席人才官 Casey。这三位员工身处完全不同的业务部门，却都在利用 ChatGPT 消化那些几个月前看起来根本无法完成的繁重工作。 Danielle 强调，人们对 ChatGPT 的认知通常停留在“头脑风暴伙伴（thought partner）”，对 Codex 的认知则是“代码编写助手（coding agent）”；但如今在企业场景中，ChatGPT 的核心定位是一个“具备代理能力的团队协作伙伴（agentic teammate）”。用户只需赋予它明确的目标意图，并输入必要的上下文背景。例如在筹备 DevDay 开发者大会时，组织者可以直接把大会日程表交给它，让它快速构建一个交互式引导页面，根据参会者当前的研发方向推荐合适的分会场。它不仅能处理单次项目，更能长期接管持续性的例行任务，把人类员工从日常琐碎的瓶颈中彻底解放出来，去投身更具创造力与探索价值的事业。 ### [03:51](https://www.youtube.com/watch?v=SXCfs7mitPo&t=231s) Dani 的实践：用智能体应对产品发布与社交舆情负责 OpenAI 开发者社交业务的 Dani，日常的核心职责是护航各类产品发布、倾听来自开发者社群的实时反馈，并持续向高管与关键利益相关者汇报发布动态。在过去，仅是盯盘监测和跨部门通报，就会耗掉她整整一天的时间。更关键的是，她的核心“利益相关者”往往是公司最高层——试想一下，当团队刚刚发布了前沿模型（例如 GPT-6 Astra），CEO Sam Altman 随时可能跑来问一句：“这次发布的反响怎么样？”Dani 的核心挑战，就是必须立刻从海量嘈杂的声音中提炼出最有价值的反馈。为了应对这一高压场景，Dani 并没有靠手动刷新网页，而是借助 ChatGPT 搭建了一个专供她和团队使用的实时动态发布看板。这个看板不仅能够综合汇总发布推文的实时表现，还能直接归纳开发者的情绪与评价。但实时看板若想真正发挥作用，就必须有人持续录入最新指标。Dani 将这项枯燥的例行差事完全交托给了她的专属智能体助手（dot）： ```mermaid flowchart TD accTitle: 社交媒体发布与舆情监控流程 accDescr: 展示智能体如何抓取社交与Slack信息并沉淀为分析数据供Dani决策 subgraph Ingestion["信息监控与采集"] A["社交媒体信息流"] -->|"实时读取"| B["专属智能体 (Dot)"] C["众多 Slack 发布频道"] -->|"后台扫描"| B end subgraph Processing["分析与更新"] B -->|"自动填充数据"| D["实时发布看板 (给高管/团队)"] B -->|"沉淀互动指标"| E["内容复盘表格"] B -->|"检测到紧急事项"| F["向 Dani 发送 Ping 提醒"] end subgraph Action["高价值产出"] D --> G["为 Sam Altman 等高管提供实时洞察"] E --> H["识别内容策略与爆款模式"] F --> I["介入高优讨论并反哺产品工程团队"] end ``` 这个专属智能体会像 Dani 本人一样浏览社交媒体流，把相关讨论提炼后自动添加到看板中；在发布结束后，它还会将每一条帖子的互动转化数据同步整理到内容追踪表格中，帮助团队直观复盘哪些内容有效、哪些策略失误。过去数小时漫无边际的机械刷帖，如今被浓缩为结构化的决策支持。Dani 因此得以抽身去思考更关键的问题：OpenAI 应该如何战略性地回复开发者？她将收集到的洞察反哺给产品工程团队以及开发者体验部门。此外，针对产品发布期漫天飞舞的 Slack 协作频道，智能体同样在后台默默值守，仅在出现紧急情况或需要她亲自表态时才会发出定向提醒，让她把宝贵的时间留给最热爱的事情——深入社群并展示开发者们的杰出作品。 - **核心价值转变**：从被动的“舆情信息搬运工”转变为深度的“社群策略与产品反馈制定者”。 - **先发制人的预判**：因为提前预判到了管理者会问“发布效果如何”，所以通过智能体让答案时刻处于“准备就绪”状态。 ### [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) Adrian 的实践：双人竞品研究团队如何杠杆化产出 Adrian 在 OpenAI 负责竞品研究，但他麾下的团队总共只有两个人。在一家飞速演进的人工智能企业中，来自销售、产品、GTM（Go-to-market）和战略等各个部门的同事每天都会涌来提出各种尖锐的竞对问题：“竞品发布了什么新功能？”“客户对他们的评价如何？”“这对外面的竞争格局意味着什么？”对于只有两人的小组来说，每一个深挖的问题都可能耗费数天案头调研，这意味着“决定调查什么，在客观上也等于决定放弃回答什么”。为了打破人力上限，Adrian 将自己全套的研究方法论（包括如何开展调查、如何识别潜在模式、如何交叉比对并推翻漏洞）系统性地沉淀到了一个定制插件中。这个插件必须承受住高强度的推敲和审视： - **颠覆表面假设**：在一次针对“某竞争对手收入为何激增”的调查中，业内普遍推测对方赢得了一大批超大型企业客户。Adrian 让他的智能体横向穿透财报电话会、分析师通讯和行业播客，搜集到的客观事实最终直接推翻了这一猜测。 - **挖掘 CRM 盲区**：在另一次复盘中，智能体自动复盘了过去 90 天内所有的销售沟通录音文本（sales call transcripts），成功找出了大量销售人员在 CRM 摘要中未曾记录的丢单真实归因，拼凑出完整的客户流失图谱。为了让全公司能自助获取洞察，Adrian 搭建了名为“Compete Corner”的内部竞品知识空间。这个空间不仅分门别类沉淀了高频竞品议题，还允许同事自行查验原始论据与事实来源，无需 Adrian 亲自出面答疑。更妙的是，Adrian 同样将更新 Compete Corner 的工作交给了智能体： ```mermaid flowchart LR accTitle: Adrian 的竞品研究自动化流水线 accDescr: 展示从输入竞品信号到自动化研究、审核与上线的完整闭环 subgraph Inputs["输入源"] S["Slack 频道提问"] M["手机短信转发竞品新闻"] end subgraph AgenticCore["智能体调研引擎"] Inputs --> P["智能体 (Dot) 触发研究插件"] P --> R["多源交叉检索与逻辑交叉验证"] end subgraph Verification["人机协同与发布"] R --> V["生成草稿放入工作区 (待审核)"] V --> A["Adrian 补充背景并点击批准"] A --> U["自动更新 Compete Corner 门户"] end ``` 在日常通勤途中，Adrian 甚至只需用手机发条短信，把在路上看到的竞对动态链接直接转给智能体。在他坐到工位之前，智能体就已经完成了信息检索、交叉比对和文章草拟，将状态标记为“待审核”。这一机制让两人团队拥有了极大的创新带宽——Adrian 还基于 GPT-6 Astra 开发了名为“Model Galaxy”的前沿模型交互式可视化系统，能够直观对比各大前沿模型在多维度基准测试下的性能、成本与延迟表现。演讲者借此鼓励台下开发者：审视自己日常排查客户问题或追查顽固 Bug 的路径，既然你清楚什么是优质的解答、何处该质疑推敲，就完全可以把这套思维框架教会 AI，让它分担枯燥的资料收集，把人的精力留给最终洞见。 ### [10:29](https://www.youtube.com/watch?v=SXCfs7mitPo&t=629s) Casey 的实践：知识库与高管 AI 教练平台最后登场的是 OpenAI 首席人才官 Casey。作为组织架构与人才战略的掌舵人，Casey 聚焦的核心命题是“人工智能如何重塑团队的协作范式”。Casey 及其团队深度调研了科技行业内部与外部的大量组织，系统性追踪不同形态的团队是如何逐步适应并融入 AI 工作流的。在长期且广泛的调研过程中，团队积累了浩繁的研究资料，涵盖大量的会议纪要、前沿学术论文以及行业分析报告。他们将这些零散的智力资产整合为一个统一的共享知识底座。基于这一扎实的知识资产，Casey 打造了一个面向管理者的专属领导力教练平台。该平台依托 GPT Live 技术构建，支持用户通过自然语音直接对话交流。管理者可以畅谈团队目前的运作阻滞点、业务瓶颈与协作卡点，AI 教练则会基于海量实证研究，协助其梳理出哪些工作可以安心下放给 AI，以及整个团队应当如何重塑协作节奏。而面对知识库如何持续迭代的挑战，Casey 的团队同样设计了一套高效的闭环。 ### [11:27](https://www.youtube.com/watch?v=SXCfs7mitPo&t=687s) 共享页面与知识沉淀：将个人探索转化为团队资产团队成员将日常的会议纪要和深度调研汇聚在同一个共享页面中，正是通过这种模式，分散的个人探索被成功凝聚为一份统一的组织级核心资源。为了保证产出质量，团队还会专门为页面设定指令（Page Instructions），明确整理新研究时的结构与规范；这样一来，当智能体（Agent）介入该页面协助撰写或归纳时，就能始终遵循统一的风格与标准。更重要的是，这套沉淀下来的知识库会直接反哺辅导站点（Coaching Site）：团队每补充一点新学到的认知，AI 教练便能在后续的指导与建议中即时调用。讲者在此向台下观众抛出一个发人深省的问题：“你自己擅长什么？你又该如何把这份专业知识开放出来，让更多人轻松获取？” - **共享沉淀**：将零散的会议笔记与个人调研集中于单一页面，避免信息孤岛。 - **页面指令规范**：通过设定格式与风格准则，确保智能体自动化处理时保持输出一致性。 - **知识闭环**：沉淀的内容直接作为底层上下文接入辅助工具（如 AI 教练），实现知识动态更新。 --- ### [12:14](https://www.youtube.com/watch?v=SXCfs7mitPo&t=734s) ChatGPT Space：多人与多 Agent 协同的全新工作空间结合当天主题演讲（Keynote）中正式公布的新功能，讲者现场展示了支持 Adrian 团队“Compete Corner”竞对情报站点的底层架构——ChatGPT Space。在 Space 中，ChatGPT 会将其搜集到的全网竞品调研直接同步到协作页面内；Adrian 与队友可以同时打开该页面协同作业，Adrian 对某个结论提出质疑，队友便能立刻添加评论并附上佐证资料。极其精妙的一点在于：两位成员在页面内均可各自唤起专属的对话线程就细节展开追问，这些对话保持彼此私密，但最终提炼并回填到页面上的内容则是全员实时共享的。 ```mermaid flowchart TD accTitle: ChatGPT Space 协作流程图 accDescr: 展示 ChatGPT、团队成员私有对话以及共享页面之间的协作与数据流转关系。 subgraph Inputs["调研与沉淀"] A["ChatGPT 外部研究成果"] end subgraph SharedSpace["ChatGPT Space 协作空间"] B["共享 Page (含 Page Instructions)"] C["Adrian 私有 Chat 追问"] D["队友私有 Chat 追问"] end subgraph Output["最终交付应用"] E["Compete Corner (竞对情报站点)"] end A --> B B <--> C B <--> D C -->|"提炼结论并编辑"| B D -->|"补充证据并编辑"| B B --> E ``` 正如代码仓库中使用 `agents.md` 文件为智能体定义规范一样，Space 中的页面指令同样起到了指引 Agent 遵循特定工作习惯的作用。除了富文本页面，Space 如今还支持文件存储，并即将在未来解锁幻灯片（Slides）和电子表格（Sheets）的原生生成与编辑功能。团队成员无需在各类软件间来回切换，即可在 ChatGPT 内部与多位 Agent 完成全流程协同。想要实践这种工作流，最佳切入点往往是截取一段冗长的日常对话，直接指令 ChatGPT “将正在讨论的内容整理成一个 Page”，随后与 Agent 结对润色，再邀请同事加入协同编辑。 - **私密对话与公共协同解耦**：成员保留个人与 AI 探索的私密上下文，同时在公共页面上同步沉淀结论。 - **类 Codebase 规范**：利用类似代码仓库的指令系统约束智能体行为，维持业务沉淀的一致性。 - **一体化扩展**：打破传统办公套件壁垒，将文档、文件以及即将上线的表格与幻灯片全部集成于 ChatGPT 内部。 --- ### [14:33](https://www.youtube.com/watch?v=SXCfs7mitPo&t=873s) 智能体副手 Dots：全天候云端值守与多端触达在团队协同之外，工作节奏的日新月异必然需要可靠的角色来全盘监控推进。讲者顺势引出了当天推出的另一项重磅功能——“Dots”（可爱的全天候智能体副手）。以 Dani 监控产品发布为例，以往需要人工不断刷社交平台，如今她将此重任全权交托给了自己的 Dot。Dot 在云端拥有独立的虚拟电脑和内置浏览器，能够像真人一样自主浏览社交动态流并实时收集热点，自动维护并刷新 Dani 的内容排期表格。更实用的是，用户可以将 Dot 直接拉入指定的 Slack 频道，并灵活配置其角色：它可以成为“墙上的苍蝇”（Fly on the wall）仅在后台静默观察，也可以化身为积极的讨论参与者；一旦检测到重大舆情或紧急事务，Dot 会立即在 Slack 上向 Dani 发送提醒（Ping）。这就赋予了 Dot 一项长期且持续的责任机制——“时刻追踪动态，仅在必要时介入并呼唤人类”。 ```mermaid flowchart LR accTitle: Dot 跨平台交互与工作流示意图 accDescr: 展示 Dot 如何在云端独立运行并通过多种终端渠道与人类进行交互。 subgraph Cloud["云端专属环境 (Always-on)"] Dot["用户专属 Dot (独立电脑 + 浏览器环境)"] end subgraph Automation["自主执行与监控"] Web["浏览社交媒体 / 网页"] Doc["更新工作表格 / 文档"] end subgraph Communication["全天候触达渠道"] Slack["Slack 频道监测与 @提醒"] Phone["短信 (SMS) 与语音通话"] Local["本地电脑文件访问 (桌面端)"] end Dot --> Web Dot --> Doc Dot <--> Slack Dot <--> Phone Dot <--> Local ``` Dot 最具颠覆性的特质在于其全天候在线（Always-on）以及无缝的跨终端一致性。不论 Adrian 是在通勤途中给 Dot 发短信、直接拨打电话，还是在办公桌前通过 Slack 发消息，背后的 Dot 始终是同一个拥有完整记忆与上下文的统一实体。讲者笑称自己给 Dot 起了儿时爱犬的名字“Wyatt”，并为其戴上了像戴单晶眼镜的滑稽青蛙外形饰品；另一位讲者则打趣说自己的 Dot 叫“Ginger”，灵感来源于辣妹组合（Spice Girls）而非宠物狗。讲者鼓励大家：无需在初始阶段把一切规划得尽善尽美，Dot 会在日复一日的协作中不断学习用户的习惯与偏好，逐渐进化为无可替代的数字副手（会场楼下还特设了供观众定制形象的 Dot Shop 实体体验区）。 - **独立云端运行环境**：自带云端系统与浏览器，支持自主检索与数据同步，无需占用本地算力。 - **上下文统一跨端体验**：电话、短信、Slack 与桌面应用无缝打通，确保沟通体验连贯不割裂。 - **自适应个性化**：外观与名字可深度定制，长期使用中主动学习使用者的工作习惯与决策风格。 --- ### [16:48](https://www.youtube.com/watch?v=SXCfs7mitPo&t=1008s) 交接乏味琐事：注意力即货币与工作重塑讲者展示了一个互动二维码，邀请台下观众掏出手机扫码体验。该程序会在移动端 ChatGPT 中启动，通过简短的交互并智能推断用户当下的工作瓶颈，进而主动提议“最应该首先移交出去的工作是哪一项”，并推荐 Space 或 Dots 的切入点。回顾开场所分享的团队标杆实践：Dani 借此能够将重心全面扑在开发者社区中；Adrian 能够以全新形式向全公司更大范围分发竞品情报；Casey 则无需被拽进每一场沟通会议，就能高效输出自己的专业智识。讲者动情地分享了自己的真实经历：刚加入 OpenAI 时，其本职工作是负责技术入职培训（Technical Onboarding），协助新入职的研发工程师快速熟悉基础架构；正是借助 ChatGPT，他将这部分繁复的基础工作彻底自动化，不仅没有被替代，反而借此顺理成章地转型至开发者体验（Developer Experience）领域，最终站在了 DevDay 的演讲舞台上。 > “注意力就是货币，而你拥有决定把它花在哪里的权利（Attention is currency and you get to decide where to spend it）。” 面对这种转变，人们起初或许会感到新奇甚至有些异样，因为这推翻了过往习以为常的做事逻辑；但它带来的真正价值是赋予每个人主动选择的权利。讲者向所有人发出了最终的灵魂拷问：“有哪些深爱的工作你一直渴望去做？有哪些搁置已久的想法、迟迟未探索的问题，或是想要多花时间陪伴的人？——你准备将什么交接出去？而你，又终于能够去承担什么真正重要的事情？”随后，讲者与 Dani 共同致谢，引导台下观众前往线下展区展开后续交流。 ## Notable quotes > [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) 每天清晨开启工作时，有谁的收件箱是完全清空的，内心无比笃定，清晰地知道今天每一步该做什么？ — Nick > [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) 又有谁在掀开笔记本电脑屏幕的瞬间，感觉就像有一个小丑突然从盒子里蹦了出来（jack-in-the-box），满眼都是无法招架的混乱？ — Nick > [00:25](https://www.youtube.com/watch?v=SXCfs7mitPo&t=25s) 具备代理能力的团队协作伙伴（agentic teammate） — Danielle > [03:51](https://www.youtube.com/watch?v=SXCfs7mitPo&t=231s) 这次发布的反响怎么样？ — Sam Altman > [06:38](https://www.youtube.com/watch?v=SXCfs7mitPo&t=398s) 决定调查什么，在客观上也等于决定放弃回答什么 — Adrian > [11:27](https://www.youtube.com/watch?v=SXCfs7mitPo&t=687s) 你自己擅长什么？你又该如何把这份专业知识开放出来，让更多人轻松获取？ — Nick 与 Danielle > [12:14](https://www.youtube.com/watch?v=SXCfs7mitPo&t=734s) 将正在讨论的内容整理成一个 Page — Nick 与 Danielle > [14:33](https://www.youtube.com/watch?v=SXCfs7mitPo&t=873s) 时刻追踪动态，仅在必要时介入并呼唤人类 — Nick 与 Danielle > [16:48](https://www.youtube.com/watch?v=SXCfs7mitPo&t=1008s) 注意力就是货币，而你拥有决定把它花在哪里的权利（Attention is currency and you get to decide where to spend it）。 — Nick > [16:48](https://www.youtube.com/watch?v=SXCfs7mitPo&t=1008s) 有哪些深爱的工作你一直渴望去做？有哪些搁置已久的想法、迟迟未探索的问题，或是想要多花时间陪伴的人？——你准备将什么交接出去？而你，又终于能够去承担什么真正重要的事情？ — Nick 与 Danielle ## Mentioned OpenAI, ChatGPT, DevDay 2026, Nick, Danielle, Dani, Adrian, Casey, Sam Altman, Codex, GPT-6 Astra, Dots, Slack, Compete Corner, Model Galaxy, GPT Live, ChatGPT Space, Dot Shop, Wyatt, Ginger, 辣妹组合 (Spice Girls) ## Video description See how OpenAI’s teams use ChatGPT to manage launches, research competitors, and turn expertise into tools. Take practical ideas back to your own work. [Source](https://www.youtube.com/watch?v=SXCfs7mitPo)

Source quotes

[0:25] Nick: 每天清晨开启工作时，有谁的收件箱是完全清空的，内心无比笃定，清晰地知道今天每一步该做什么？ [0:25] Nick: 又有谁在掀开笔记本电脑屏幕的瞬间，感觉就像有一个小丑突然从盒子里蹦了出来（jack-in-the-box），满眼都是无法招架的混乱？ [0:25] Danielle: 具备代理能力的团队协作伙伴（agentic teammate） [3:51] Sam Altman: 这次发布的反响怎么样？ [6:38] Adrian: 决定调查什么，在客观上也等于决定放弃回答什么 [11:27] Nick 与 Danielle: 你自己擅长什么？你又该如何把这份专业知识开放出来，让更多人轻松获取？ [12:14] Nick 与 Danielle: 将正在讨论的内容整理成一个 Page [14:33] Nick 与 Danielle: 时刻追踪动态，仅在必要时介入并呼唤人类 [16:48] Nick: 注意力就是货币，而你拥有决定把它花在哪里的权利（Attention is currency and you get to decide where to spend it）。 [16:48] Nick 与 Danielle: 有哪些深爱的工作你一直渴望去做？有哪些搁置已久的想法、迟迟未探索的问题，或是想要多花时间陪伴的人？——你准备将什么交接出去？而你，又终于能够去承担什么真正重要的事情？

Source quotes

[0:25] Nick: like a clown has jumped out of your laptop [0:25] Nick: like a jack in the box [0:25] Nick: when things slow down [11:27] What are you an expert on? And how could you make that accessible to more people? [14:21] a fly on the wall [14:21] tell it what you want, what you really, really want [16:45] Nick: a little bit strange [16:45] Nick: Attention is currency and you get to decide where to spend it. So, let me ask you: What will you hand off? And what could you finally take on?

Original transcript

[0:25] Hello. [0:26] Thank you for joining us today. [0:27] My name is Nick. I'm Danielle. [0:29] And we work on the developer [0:31] experience team here at OpenAI. [0:33] Let me start with a poll [0:35] from the crowd. [0:36] Go ahead and raise your hand if when [0:38] you start your day, your inbox is [0:40] empty, you're full of clarity, and you [0:42] know exactly what you need to do. [0:44] You've got one, two. [0:46] All right. Let's put your hands down. [0:48] All right. Now, go ahead and raise your hand [0:50] if when you open your laptop, [0:51] it feels like a clown has [0:53] jumped out of your laptop. [0:54] It's full of chaos. [0:56] It's like a jack in the box. [0:57] All right. [0:58] A lot more hands. [1:00] All right. [1:01] So here's what that looks like. [1:03] You open your laptop, you check [1:05] your email, you catch up on Slack, [1:07] and a couple dashboards. [1:09] And then you finally [1:11] decide we need to get done. [1:12] But the problem is, it's everything. [1:15] And by now, half the morning is gone, [1:17] and you're just starting your real job. [1:20] So we know that ChatGPT can change this. [1:23] But how? [1:26] So these are our most prolific [1:28] ChatGPT work users here at OpenAI. [1:31] They're all outliers within [1:33] their respective orgs. [1:34] And me and Danielle interviewed them [1:36] because we wanted to know exactly [1:38] what they were doing differently. [1:39] What was their secret? [1:41] So we're going to show you exactly [1:43] how they're leveraging ChatGPT and what [1:45] else that empowers them to do. [1:49] So let me introduce our champions. [1:52] This is Dani. [1:53] She leads social for OpenAI developers. [1:57] We then have Adrian, [1:59] who leads competitive research. [2:01] And finally, Casey, who's our chief [2:04] people officer here at OpenAI. [2:07] These are three very, [2:08] very different jobs. [2:10] But our conversations with Dani, [2:12] Adrian, and Casey revealed [2:15] a pattern in how they had [2:17] taken on work that seemed [2:19] impossible just a few months ago. [2:22] Each of them is a ChatGPT [2:25] work power user. [2:26] You already know ChatGPT [2:28] as a thought partner. [2:30] And you know Codex as a coding agent. [2:33] ChatGPT work is your agentic [2:35] teammate for getting things done. [2:38] You tell it what you want done, [2:40] and you give it the context [2:41] that it needs to do it. [2:45] Let's say I'm running [2:47] an event like DevDay. [2:49] I can hand over the agenda [2:51] for DevDay and have it build [2:53] a site that asks what you're [2:55] working on and helps you pick [2:57] which sessions you should attend. [2:59] Totally unbiased, of course. [3:05] ChatGPT work, or as we usually call it, [3:07] chat, can also take on ongoing tasks. [3:10] And when you can hand off [3:12] more of your day-to-day work, [3:14] you can spend more time working on [3:16] what you're truly excited about. [3:20] You probably have [3:21] something in mind already. [3:23] Maybe it's an idea that you [3:24] really want to explore. [3:27] Maybe it's a project that you'd [3:28] love to get started on. [3:30] And you will, you know, [3:32] when things slow down. [3:34] What's in the way? [3:37] Where are you the bottleneck? [3:39] And if Chat could take on [3:40] a responsibility for you, what [3:43] would that free you up to do? [3:45] So let's see what that handoff [3:47] enabled for Danny, Adrian, and Casey. [3:51] Let's start with Danny. [3:53] Danny leads developer [3:54] social here at OpenAI. [3:56] So her job is to shepherd products [3:58] through launch and then listen [3:59] to the developer community for feedback. [4:01] She also keeps key stakeholders in loop [4:04] with how those launches are going. [4:05] And Danny told us that she [4:07] could spend an entire day [4:08] just monitoring these launches [4:10] and then keeping people in the loop. [4:12] And for Danny, [4:13] stakeholders mean executives. [4:15] So imagine you just helped [4:17] launch GPT-6 Astra. [4:18] And now the CEO, Sam Altman, is [4:21] asking you how the launch is going. [4:23] And so Danny's job is to distill [4:26] the feedback from the noise. [4:27] And here's how she does it. [4:30] So she uses ChatGPT to pull together [4:33] how the launch posts are performing [4:34] and what developers are saying. [4:36] And she pulls that into a ChatGPT [4:38] site for her and her team. [4:40] So when Sam is asking, you know, [4:42] how's the launch going, she [4:44] can share a live report. [4:46] But a live report is only [4:48] useful if somebody is keeping [4:49] up to date and keeping those [4:50] metric numbers up to date. [4:53] And so that's the job that she's [4:55] handed off to her dot. [4:58] It reads the social feed just like [5:00] Danny would and adds posts to her site. [5:03] After the launch, she can share [5:04] a more detailed view with her [5:06] and her team because that same [5:07] dot has been adding posts to her [5:09] content sheet with engagement [5:10] metrics for every single post. [5:13] This makes it easier for her [5:14] and her team to understand what content [5:16] is working and what's not working. [5:18] So Danny used to do most [5:20] of that monitoring and collection [5:21] herself, which meant hours of scrolling. [5:24] But now she can focus [5:26] on the bigger picture, [5:28] like figuring out how we should respond [5:29] to those developers. [5:30] And also, this is just better than [5:32] only collecting the sentiment. [5:34] Also, she can bring back what [5:36] she's learning to people in product [5:37] engineering, as well as, you know, very [5:40] important stakeholders like Danielle [5:42] and myself in developer experience. [5:46] She also has to keep up [5:48] with everything happening in Slack. [5:50] The sheer number of channels [5:52] that Danny has to be in because [5:53] she's helping launch so many products [5:55] means that it's really easy to not [5:57] keep up with all those updates. [5:59] So what she's done is she has [6:01] her dot monitor those channels for her. [6:03] It pings her when something's [6:05] urgent and lets her know [6:06] when she needs to chime in. [6:09] All of this means that Danielle [6:11] can spend more time with the developer [6:14] community and spotlight their work. [6:16] And this is the work that Danny [6:17] loves to do. [6:23] So, Danny can hand off all [6:25] of that work because she knows what [6:27] questions she's going to get asked. [6:29] So, think about what question [6:31] you're going to get asked [6:32] and how could chat have the answer [6:35] ready before somebody asks. [6:38] And that leads us to Adrian. [6:40] So, Adrian leads competitive research. [6:42] His team is just two people. [6:45] Colleagues from sales, [6:46] from product, GTM, strategy, [6:49] all across the org come to Adrian [6:51] with questions about the competition. [6:53] What's changed [6:54] in a competitor's product? [6:56] What are customers saying? [6:58] And what does that mean for us? [7:01] Each question can turn into days [7:03] of research and with a small [7:05] team of just two. [7:07] Deciding what to investigate also [7:10] means deciding what goes unanswered. [7:14] So, to answer more of those questions, [7:17] Adrian gave chat a way to take [7:19] on more of that research. [7:20] He put his full research [7:22] method into a plugin. [7:24] And that plugin does what he does. [7:27] It investigates, it finds patterns, [7:30] and it challenges the findings. [7:32] He needs an answer that can [7:34] hold up well under scrutiny. [7:36] So, how does that actually work? [7:41] Adrian wanted to understand [7:42] why a competitor's revenue [7:44] was growing so quickly. [7:46] Was it winning a new group [7:48] of really large customers? [7:49] He had his DOT look across [7:52] calls, across newsletters, [7:54] and podcasts. [7:56] The evidence didn't [7:57] support that explanation. [8:00] So, in another investigation, [8:02] he had his DOT go back through [8:05] 90 days of sales call transcripts [8:08] and found reasons for lost deals [8:10] that were not captured in the CRM [8:13] summaries that the salespeople left. [8:15] And that gave him a much [8:16] fuller picture of why customers [8:18] were choosing a competitor. [8:21] He also built Compete Corner. [8:24] It's a site that organizes research [8:27] around the questions that he [8:28] knows he's going to get asked. [8:30] Colleagues can go to Compete Corner [8:32] and explore findings and check [8:34] the sources without asking [8:37] Adrian to walk them through it. [8:39] Adrian passed off the job of keeping [8:42] Compete Corner up to date. [8:44] His DOT picks up the competitive [8:46] questions straight from Slack. [8:48] It then researches [8:49] them using his plugin. [8:51] And then it brings those findings [8:53] into his space for review. [8:56] Adrian or his teammate [8:57] can then check the conclusions [8:59] and add additional context. [9:01] Once they approve, his DOT can then [9:03] update Compete Corner on his behalf. [9:08] Adrian can also just [9:10] send his DOT a text. [9:12] On his commute in the morning, [9:14] he sees a competitor announcement. [9:16] He can forward it straight to his [9:18] DOT, where it does exactly the same [9:20] workflow without Adrian needing to be [9:22] in ChatGPT or at his laptop at all. [9:25] And the page is marked ready for review [9:28] when he arrives at his desk. [9:30] So what does that change [9:32] for a team of two like Adrian's? [9:34] Well, it enables Adrian to do [9:37] all sorts of new things like [9:39] delivering competitive intel [9:41] in an entirely new way. [9:43] He built this tool. [9:44] It's called Model Galaxy. [9:45] And it's an interactive [9:47] visualization of performance, [9:49] cost, and latency across any [9:52] benchmark for frontier models. [9:54] Of course, made with GPT-6 Astra. [9:58] So think about an investigation [10:00] that you do regularly. [10:02] Maybe it's a customer problem. [10:04] Maybe it's a bug that just [10:05] refuses to be squashed. [10:07] You already know what makes for a good [10:09] answer and when to challenge it. [10:12] You know what sources you should trust. [10:14] So show Chat how you would approach [10:16] that exact same investigation. [10:19] It can do more of the gathering [10:21] and the checking so that you can [10:23] actually focus on the findings [10:25] and what you should do about them. [10:29] What if people could draw on [10:31] your expertise without having to come [10:33] to you every time they have a question? [10:35] This brings us to Casey. [10:36] He's our chief people [10:37] officer here at OpenAI. [10:39] And he's thinking about how AI is [10:41] changing the way teams work together. [10:44] So Casey and his team [10:45] study organizations inside [10:47] and outside the industry. [10:48] And they're trying to learn how these [10:50] teams are adapting to AI. [10:52] And through all this research, [10:54] they are accumulating a treasure trove [10:56] of meeting notes, papers, and articles. [10:59] And they bring all this research [11:00] together into one shared knowledge base. [11:04] He's used that research to build [11:06] a coaching site for leaders. [11:08] It's powered by GPT Live. [11:10] So you can talk to it about how [11:11] your team works and where [11:12] things are getting stuck. [11:14] It helps you figure out what to hand [11:15] over to AI and how your team [11:17] could work differently. [11:18] But how does Casey's team [11:20] keep this knowledge base that's [11:22] powering the coach up to date? [11:25] So Casey and his team bring [11:27] their meeting notes and research [11:29] together on a shared page. [11:30] And this is how they leverage [11:32] their individual explorations [11:33] together into one resource. [11:35] They also give the page instructions [11:38] for organizing new research. [11:40] So when an agent works in that page, [11:42] it can ascribe to a consistent style. [11:45] And this same knowledge [11:47] feeds the coaching site. [11:48] As the team adds what they're learning, [11:51] the coach can use it in his suggestions. [11:54] So go ahead and ask yourself, [11:56] what are you an expert on? [11:58] And how could you make [11:59] that accessible to more people? [12:04] So you've now seen what our champions [12:07] are taking on with chat. [12:09] And now we're going to show [12:10] you how to put it to work. [12:14] As you've probably already [12:15] seen in the keynote today, [12:17] we've launched ChatGPT space. [12:21] Yeah, that's awesome. [12:25] So you've already seen [12:27] Adrian's Compete Corner site. [12:28] Now I'm going to show you [12:30] the space that hydrates it. [12:33] Chat brings its research [12:35] into Adrian's space. [12:37] He and his teammate work through [12:38] the findings, all in the same [12:40] page at the same time. [12:41] Adrian can question a conclusion. [12:43] His teammate can comment and add [12:46] a source that helps answer it. [12:47] They each work with their own [12:49] chat in the space [12:50] to investigate further right there. [12:52] Their chats each stay private to them, [12:55] but the edits on the page are shared. [12:58] The research that they develop [12:59] here then feeds Compete Corner. [13:05] We also mentioned that Casey [13:08] is using page instructions [13:09] to keep things consistent. [13:11] Like an agents .md file [13:14] for a code base, they describe [13:16] the working practices that agents [13:18] should follow on a particular page. [13:21] And pages are just the beginning. [13:24] You can also store files [13:26] in your space and soon you're [13:27] going to be able to create slides, [13:29] sheets and so much more. [13:30] So you and your teammates [13:32] and your agents work together [13:35] all in the same place [13:36] without leaving ChatGPT. [13:41] So to try this yourself, [13:43] bring the work that you want [13:45] to do with Chat into your space. [13:49] Add the context that it needs [13:50] and then you can invite your teammates. [13:52] My favorite way to do this is [13:55] to start with a long Chat thread [13:57] and ask it to turn what you're [13:58] working on into a page. [14:01] From there, you can pair directly [14:03] with your agent and continue refining. [14:05] As you talk to your agent, [14:07] it's going to keep the page updated. [14:10] Then, you can invite your teammates. [14:12] They can edit the same [14:14] page at the same time. [14:16] And again, those edits [14:17] become shared in the page. [14:21] Now, working together in one [14:23] space is a great way to start. [14:26] But work continues to change [14:28] and somebody does have [14:29] to keep track of it all. [14:33] You probably know what's coming next [14:35] because today we have [14:36] also launched Dots. [14:38] Oh, so cute. [14:41] So remember how Dani handed [14:43] off monitoring the launch? [14:45] So here's how she did it. [14:46] Her Dot has its own [14:48] computer in the cloud as [14:49] well as its own browser. [14:51] So it goes to the social feed [14:52] just like Dani would and it [14:54] collects the posts while keeping [14:55] her content sheet up to date. [14:58] Then, she can add it to the Slack [14:59] channel she wants it to follow. [15:01] And she can also tell what to do. [15:03] This could be just be a fly [15:04] on the wall or also be [15:05] an active participant. [15:07] So when something needs her attention, [15:09] it can send her a ping on Slack. [15:11] So what she's done here is she's [15:13] given her Dot an ongoing responsibility. [15:15] Keep track of what's happening and then [15:17] bring her in when she's needed. [15:22] And you can hand off something [15:23] from wherever you are. [15:24] We mentioned that Adrian can [15:26] text his Dot in his commute. [15:28] He can also just give his Dot [15:29] a call or ping on Slack. [15:31] His Dot is reachable wherever he is [15:33] and it's always the same [15:34] Dot with the same context. [15:36] So when he's messaging on Slack [15:37] or he's getting a call, [15:38] it's always the same entity. [15:40] And your Dot is always on [15:41] because it lives in the cloud. [15:43] It can text you an update or a [15:45] question when it needs you to weigh in. [15:46] And it can also access your local [15:48] files when you're on your laptop. [15:51] So how do you turn your Dot [15:52] into your own sidekick? [15:55] Well, the first thing you're going to do [15:56] is you're going to give it a name. [15:58] I've named mine Wyatt. [15:59] It's the name of my childhood dog. [16:02] The next thing you're going to do [16:03] is give accessories. [16:05] I like having mine kind of look [16:06] like the frog guy with the fun [16:07] eyes and the monocle. [16:09] And after that, you're going [16:10] to give it a responsibility. [16:11] You're going to tell it what you want, [16:12] what you really, really want. [16:15] And then you're going to tell [16:15] it where to reach you. [16:16] You can add it to Slack. [16:17] You can have it text you when [16:19] messaging is coming out soon. [16:21] But the point is you don't have [16:23] to figure this all out up front. [16:25] Your Dot will kind of learn [16:27] your working style as you work with it. [16:28] It'll learn your preferences and it'll [16:30] get better over time. [16:32] By the way, if you would like [16:34] to build your own custom Dot, [16:35] we do have a Dot shop downstairs. [16:37] So do check that out after this talk. [16:40] I named my Dot Ginger after [16:42] the Spice Girl, not a dog. [16:45] All right. And now it's time for you [16:48] to hand off the part [16:49] of your job that you won't miss. [16:50] So I want you to feel free [16:52] to take a picture of this QR code. [16:53] You can do it right now. [16:54] Pull your phone out. [16:55] This is going to open up ChatGPT [16:57] on your phone. [16:58] And it will ask you [16:59] just a little bit about your work. [17:00] It'll also try to infer your work. [17:03] And once it does that, it's [17:04] going to suggest something [17:05] that you can hand off first. [17:07] And it will also tell you where [17:09] space and dots might fit into that. [17:12] So feel free to take a picture. [17:14] Hold on to that for later. [17:16] We still have a little more. [17:17] So, you know, come back [17:18] to that in a second. [17:22] A few more. [17:24] A few more photos. [17:27] All right. [17:28] All right. Moving on. [17:30] So we started this by asking what [17:32] are champions we're doing differently? [17:34] And now you're in on their secret. [17:37] Danny can spend more time [17:38] with the developer community. [17:40] Adrian can deliver competitive insights [17:42] to more people and in new ways. [17:45] And Casey can share his [17:46] expertise without having [17:48] to be in every conversation. [17:51] I have a personal version of this. [17:55] I started at OpenAI [17:58] working on technical onboarding. [18:00] My job was to help new engineering [18:03] colleagues onboard and get up [18:05] to date on our infrastructure [18:06] and our architecture. [18:08] I automated a lot of that work [18:11] away with ChatGPT. [18:13] It's a true story. [18:15] And that freed me up to move [18:16] to developer experience. [18:18] And now I'm here giving this talk. [18:22] Yeah. [18:23] Thanks. [18:25] Now that felt exciting, but it [18:27] also felt a little bit strange. [18:29] I was changing work that I [18:30] knew how to do and making [18:32] room for work that I wanted [18:34] to and was excited to try. [18:36] But that's what I [18:37] want to leave you with. [18:38] A choice. [18:40] About where more of your day goes. [18:44] You know the work that you love doing. [18:46] Maybe there's an idea [18:48] you keep putting off. [18:49] A question that you [18:51] finally get to explore. [18:53] Or maybe you just want to spend [18:55] more time with the people [18:57] that you want to help. [18:59] Attention is currency and you [19:00] get to decide where to spend it. [19:05] So, let me ask you. [19:07] What will you hand off? [19:10] And what could you finally take on? [19:14] Thank you. [19:20] All right. [19:21] Thank you all for joining us. [19:22] If you guys have any questions, come [19:24] find me and Dani after the talk. [19:25] Thanks, guys. [19:28] Thanks, [19:32] guys.