OpenAI Dots Finds Missed Ad Revenue but Stumbles on Multistep Workflows

Peter Yang

OpenAI's Dots succeeds as an autonomous orchestrator that can audit back-office revenue, but ChatGPT's cluttered interface and silent tool failures prevent it from managing complex business operations without constant human oversight.

AI-assisted summary of ChatGPT's NEW Dots: Everything You Need To Know — original source. Check the original for full context.

Source published: 2026-09-29

Key ideas

OpenAI's new 'dots' agent lives natively inside ChatGPT for Pro subscribers, operating across both a 24/7 remote cloud instance and a local Mac connection to handle ongoing personal and business tasks. (0:42)
Functioning like an executive Chief of Staff, dots coordinates work across specialized sub-threads and autonomously uncovered unbilled sponsorship spots alongside a $10,000 Google billing discrepancy during a revenue audit. (9:28)
The agent excels at deep diagnostic research—such as synthesizing 1,800 YouTube comments across 300+ videos in 30 minutes and auditing Slack communities—but still produces 'borderline slop' copy when asked to one-shot a $99 digital product. (7:38)
Because dots manages numerous sub-threads inside a single chat log, critical authentication and tool failures easily get buried in a wall of text, highlighting growing UI sprawl across ChatGPT. (13:29)

Detailed notes

## Background Published by Peter Yang on YouTube on September 29, 2026, this video tests OpenAI's new always-on agent, "dots," integrated into ChatGPT. According to the description, Yang evaluates dots across his solo business operations—such as catching missed sponsorship revenue, preparing a fitness app launch, and building landing pages—while contrasting its performance with competitors Grok Bot and Muse.

## TL;DR Peter Yang demonstrates how OpenAI's dots operates as an executive "Chief of Staff" agent within ChatGPT, orchestrating sub-threads and tapping into native memory across cloud and local Mac environments. Through practical experiments, the agent successfully flagged an unbilled five-figure sponsorship discrepancy, audited Slack communities, and analyzed over 1,800 YouTube comments to extract strategic growth levers. However, it struggled with tool authentication failures, produced substandard marketing copy for a paid product, and highlighted ChatGPT's increasingly cluttered user interface. Ultimately, Yang views dots as a powerful single-player business orchestrator that points toward a future AI-native collaboration platform competing directly with Grok Bot and Slack.

## Atmosphere & style The video maintains a pragmatic, creator-first tech review style that feels grounded in real-world business stakes rather than speculative hype. Yang speaks with measured candor, blending analytical product management critiques with genuine surprise when the agent catches real financial errors. The demonstration carries an experimental, live-testing rhythm that alternates between fluid productivity wins and the mild frustration of navigating tool failures and user-interface clutter.

## Bulletin - [00:42](https://www.youtube.com/watch?v=z5X1eMU6isI&t=42s) OpenAI's new 'dots' agent lives natively inside ChatGPT for Pro subscribers, operating across both a 24/7 remote cloud instance and a local Mac connection to handle ongoing personal and business tasks. - [09:28](https://www.youtube.com/watch?v=z5X1eMU6isI&t=568s) Functioning like an executive Chief of Staff, dots coordinates work across specialized sub-threads and autonomously uncovered unbilled sponsorship spots alongside a $10,000 Google billing discrepancy during a revenue audit. - [07:38](https://www.youtube.com/watch?v=z5X1eMU6isI&t=458s) The agent excels at deep diagnostic research—such as synthesizing 1,800 YouTube comments across 300+ videos in 30 minutes and auditing Slack communities—but still produces 'borderline slop' copy when asked to one-shot a $99 digital product. - [13:29](https://www.youtube.com/watch?v=z5X1eMU6isI&t=809s) Because dots manages numerous sub-threads inside a single chat log, critical authentication and tool failures easily get buried in a wall of text, highlighting growing UI sprawl across ChatGPT.

## Key takeaways - [00:42](https://www.youtube.com/watch?v=z5X1eMU6isI&t=42s) Dots is an always-on autonomous agent pinned to the top of the ChatGPT desktop sidebar that supports voice calling, one-click Slack integration, and dual execution across a 24/7 cloud computer and a local Mac. - [01:34](https://www.youtube.com/watch?v=z5X1eMU6isI&t=94s) Unlike consumer concierge agents like Muse that make outbound phone calls to physical businesses, dots competes directly with Grok Bot on knowledge work and business operations. - [01:34](https://www.youtube.com/watch?v=z5X1eMU6isI&t=94s) Dots' biggest advantage over Grok Bot is zero-setup onboarding: it automatically inherits all existing ChatGPT conversation history, custom instructions, memory, and authenticated tools. - [03:18](https://www.youtube.com/watch?v=z5X1eMU6isI&t=198s) Connected to Slack in one click, dots audited an entire community archive in seven minutes, grading it an 'A-minus' and proposing three concrete initiatives to reach an 'A', while also optimizing a Maven course landing page. - [03:18](https://www.youtube.com/watch?v=z5X1eMU6isI&t=198s) When generating memes in Slack, dots excelled at hyper-contextual inside jokes based on user chat history, though its generic AI industry humor fell flat. - [07:38](https://www.youtube.com/watch?v=z5X1eMU6isI&t=458s) Tasked with analyzing six months of YouTube feedback, dots spent 30 minutes parsing over 1,800 comments across 300+ videos, concluding that viewers most want transparent API costs, blank-setup tutorials, and post-demo production realities. - [08:29](https://www.youtube.com/watch?v=z5X1eMU6isI&t=509s) Dots acts as an orchestrator rather than a single-thread executor, spinning up sub-threads to generate marketing layouts, fix code bugs, capture live app screenshots, and assemble App Store assets for a fitness app launch. - [09:28](https://www.youtube.com/watch?v=z5X1eMU6isI&t=568s) Because dots retains full ChatGPT memory, prompting it to suggest ambitious initiatives yielded five tailored business projects, including an automated sponsorship revenue audit. - [09:28](https://www.youtube.com/watch?v=z5X1eMU6isI&t=568s) During the sponsorship audit, dots acted as an automated revenue manager, catching unbilled Riverside ad spots and flagging an unresolved $10,000 billing mismatch with Google. - [11:26](https://www.youtube.com/watch?v=z5X1eMU6isI&t=686s) When asked to build a $99 'PM Workbench' digital product from scratch, dots generated polished visuals but wrote generic, 'borderline slop' product copy that required heavy human revision. - [11:26](https://www.youtube.com/watch?v=z5X1eMU6isI&t=686s) Using its computer-use browser agent, dots navigated live Substack analytics dashboards and synthesized a five-part tactical growth strategy focused on paid funnels and cross-recommendations. - [13:29](https://www.youtube.com/watch?v=z5X1eMU6isI&t=809s) Complex multi-step workflows across external tools like Riverside and Substack can stall due to authentication errors, and those failures get buried inside dots' noisy, single-stream orchestration log. - [13:29](https://www.youtube.com/watch?v=z5X1eMU6isI&t=809s) ChatGPT is suffering from severe UI and feature sprawl—accumulating standard chats, Codex, Spaces, scheduled tasks, and dots—which creates cognitive overhead that makes voice mode a cleaner way to interact with the agent. - [16:17](https://www.youtube.com/watch?v=z5X1eMU6isI&t=977s) While Grok Bot already supports multiplayer team bot swarms, dots is currently a single-player experience, pointing toward a larger industry race between OpenAI, xAI, and Anthropic to build an AI-native alternative to Slack.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=z5X1eMU6isI&t=0s) Overview: Testing OpenAI's "Dots" Agent

OpenAI has introduced "dots", an always-on AI agent built directly into ChatGPT aimed at handling ambitious personal and business tasks. Over several days of testing it to run a solo operation, the agent uncovered a five-figure sponsorship deal that had slipped through the cracks, identified multiple bugs inside an in-progress fitness application, generated ready-to-use App Store marketing screenshots, pitched a new paid product concept with a generated landing page, and even joined Slack to trade humorous memes.

Across these experiments, the agent demonstrated clear utility for business operations alongside several rough edges. Evaluating whether dots outperforms rival autonomous assistants like Grok Bot and Muse comes down to distinct product architectures, feature sets, and target audiences.

```mermaid flowchart TD accTitle: Dots Operational Architecture accDescr: Visual mapping of how Dots connects local and cloud environments into ChatGPT.

subgraph ChatGPTHub["ChatGPT Native Environment"] Dot["Dot Agent (e.g., 'Zoe')"] Memory["ChatGPT Native Memory & Chat History"] Voice["ChatGPT Voice Calling"] end

subgraph Integrations["External Access Points"] Slack["Slack Integration"] Email["Email (Coming Soon)"] end

subgraph Execution["Dual Execution Environments"] Cloud["24/7 Remote Cloud Computer"] Local["Local Mac Laptop / Browser"] end

Dot <--> Memory Dot <--> Voice Dot <--> Slack Dot -.->|Future| Email Dot <--> Cloud Dot <--> Local ```

[00:42](https://www.youtube.com/watch?v=z5X1eMU6isI&t=42s) How Dots Operates Inside ChatGPT

Within the revamped ChatGPT desktop interface, dots appears as a dedicated pinned conversation at the very top of the sidebar. Users can personalize their dot's persona; for this setup, the agent is named "Zoe" and configured with a custom avatar displaying a triangle and apples.

The agent's profile provides several direct communication pathways and system connectors: - **Calling**: Direct audio conversations powered by ChatGPT Voice. - **Messaging Platforms**: Native one-click workspace access via Slack. - **Email Gateway**: A dedicated email hook (present in the interface, though marked as currently unavailable). - **Dual-Machine Environment**: The agent operates across two computing environments simultaneously—a remote 24/7 cloud instance for persistent tasks, and a local machine connection running alongside ChatGPT on a Mac laptop. - **Activity Feed**: A built-in dashboard detailing real-time logs of queued and executed assignments.

[01:34](https://www.youtube.com/watch?v=z5X1eMU6isI&t=94s) Dots vs. Grok Bot vs. Muse

Dots does not directly compete with Muse, because Muse focuses almost entirely on personal consumer logistics—such as placing outbound phone calls to physical American businesses, booking flights, or negotiating items on Facebook Marketplace (for example, successfully calling a bakery to order a wife's birthday cake). Instead, dots competes directly with Grok Bot on knowledge work, deep research, and business workflows.

Dots' biggest advantage over Grok Bot is seamless integration with ChatGPT. Because Zoe automatically inherits all existing conversation logs, customized instructions, and authenticated accounts, onboarding requires zero credential setups or account linking.

[03:18](https://www.youtube.com/watch?v=z5X1eMU6isI&t=198s) Social Research on X, Slack Audits, and Internal Memes

For social media intelligence, Zoe was prompted to scout trending practical AI posts on X and summarize them directly in chat—mirroring an identical task assigned to Grok Bot's dedicated "X Scout" agent. Running via local browser automation on a Mac, both agents returned solid selections, though they surfaced entirely distinct sets of tweets.

Moving the workflow to Slack demonstrated deeper operational analysis. Connecting Zoe to a community channel took just one click, where the agent was placed into a direct-message thread with an AI instructor named Colin, creator of an educational course on Maven: - **Course Landing Page Optimization**: Zoe spent six minutes analyzing competitive Maven courses, returning a proposal headline urging Colin to "sell the finished work," along with three strategic messaging revisions. - **Community Health Audit**: Asked to review every message thread in the Slack group and assign a grade, Zoe scanned the archive in seven minutes. It awarded an "A-minus" and outlined three clear initiatives to reach an "A": hosting monthly live sessions with 5-to-10-minute member demos followed by breakouts, building a searchable knowledge portal for durable lessons, and ensuring every single member help request receives a defined next step.

When tasked with generating five community memes featuring Peter and Colin based on their chat history ("Don't disappoint me"), Zoe excelled at hyper-contextual humor—mocking Peter's obsession with bleeding-edge AI tools versus Colin's measured, principled approach. However, broader prompts regarding generic AI industry headlines (such as Claude 3.5 Sonnet beating *Pokémon Red*) or DevDay predictions fell flat. OpenAI employees maintain an internal Slack channel exclusively populated by their own dots, where autonomous agents trade memes back and forth continuously throughout the day.

[07:38](https://www.youtube.com/watch?v=z5X1eMU6isI&t=458s) Analyzing 1,800 YouTube Comments

To stress-test deep data analysis, Zoe was instructed to read through every comment left on Peter's YouTube channel over the preceding six months and surface the top five growth recommendations. Gathering the raw data required parsing feedback across more than 300 videos, requiring roughly 30 minutes of processing time to synthesize over 1,800 comments.

The agent concluded that the single highest-leverage growth lever was making complex technical workflows easier to replicate and trust: - Clearly disclose the true operational and API costs required to run each workflow. - Ensure every tutorial can be followed end-to-end from an absolute blank setup. - Reveal "the part after an impressive demo"—showing what ongoing maintenance, edge cases, and actual production usage look like.

[08:29](https://www.youtube.com/watch?v=z5X1eMU6isI&t=509s) Fitness App Launch Prep and Sub-Thread Delegation

Rather than trying to write code and generate media inside a single prompt, dots coordinates work across dedicated threads. When assigned to review Peter's fitness app build thread, understand the code base, and design six launch-ready App Store promotional assets, Zoe acted as an orchestrator.

The agent spawned a sub-thread to prompt ChatGPT's image generator for marketing layouts, and forwarded diagnostic prompts into the core build thread to find and resolve bugs in the codebase. By running live tests of the application, Zoe captured authentic interface screenshots, layered clean promotional headlines on top, and prepared the complete visual suite needed for the App Store submission process.

[09:28](https://www.youtube.com/watch?v=z5X1eMU6isI&t=568s) Dots as Chief of Staff: Catching a Five-Figure Sponsorship Issue

Dots is built primarily as an orchestrator thread rather than a direct execution tool. Instead of asking dots itself to perform low-level tasks, you delegate to it so it can spin up, manage, and coordinate other ChatGPT conversation threads, gathering status updates as those sub-threads make progress. Functionally, it operates like an executive Chief of Staff or a project manager supervising subordinate worker threads.

```mermaid flowchart TD accTitle: Dots Orchestration Workflow accDescr: Diagram illustrating how the user issues commands to dots, which then manages multiple specialized ChatGPT execution threads and reports results back. User["User"] -->|"Issues commands"| Dots["dots (Chief of Staff)"] subgraph Execution ["Managed ChatGPT Threads"] Audit["Sponsorship Audit"] Product["Product Builder"] Substack["Substack Growth"] end Dots --> Audit Dots --> Product Dots --> Substack Audit -.->|"Discrepancies found"| Dots Product -.->|"Site draft ready"| Dots Substack -.->|"5-step growth plan"| Dots ```

A powerful approach to unlock its capability is prompting it directly: *"Let's try to do some ambitious things. What else can you do? Give me your top five."* Because dots lives inside ChatGPT and retains memory across historical conversations, it generated five tailored, high-leverage initiatives based on existing business context: 1. Build a new paid product addressing audience pain points. 2. Shepherd a video project from initial concept to publish-ready state. 3. Execute a comprehensive sponsorship revenue audit. 4. Run the fitness app through an end-to-end launch rehearsal. 5. Transform the community offering into an active, hands-on experience.

The sponsorship audit immediately proved its financial value for a solo business owner whose primary livelihood relies on ad deals. Dots autonomously audited historical sponsor logs and surfaced multiple billing oversights: - **Riverside invoicing:** Discovered that Riverside had not been billed for completed sponsorship spots, prompting an immediate invoice. - **Google discrepancies:** Flagged an unresolved $10,000 billing mismatch on Google sponsorship campaigns. - Proactively acting as an automated revenue manager holding the creator accountable without requiring manual spreadsheet audits.

---

[11:26](https://www.youtube.com/watch?v=z5X1eMU6isI&t=686s) Prototyping Paid Products and Substack Growth Analysis

Following through on its own suggestions, dots attempted to construct a paid product from scratch aimed at product managers. It generated a ChatGPT-hosted website concept dubbed the "PM Workbench," positioned as a $99 digital workflow kit featuring four interactive sample artifacts: an evidence brief, a competitive scan, a stakeholder update, and a PRD red team evaluation.

While the visual assets generated via ChatGPT looked polished, the product copy fell short. As a product manager with a decade of industry experience, the output came across as "borderline slop" rather than credible, market-ready material, demonstrating that dots cannot yet one-shot a viable digital business asset without heavy human revision.

In contrast, its diagnostic research capabilities proved far stronger. When asked to evaluate Substack newsletter metrics, dots launched a computer-use browser agent, navigated the creator's live Substack statistics dashboards, collected the relevant numbers, and extracted a five-part tactical growth strategy: - Overhauling the paid membership purchase funnel. - Aggressively scaling newsletter cross-recommendations. - Streamlining discovery and subscription conversion loops.

---

[13:29](https://www.youtube.com/watch?v=z5X1eMU6isI&t=809s) Tool Failures, Cluttered Threads, and the UI Sprawl Dilemma

Dots reveals clear functional friction when executing complex, multistep automations across external integrations. During an attempt to process a Riverside interview transcript with Mickey—using custom skills like podcast prep and podcast editing to produce a Substack newsletter draft—the workflow stalled due to authentication and tool-availability errors in Substack.

Crucially, dots failed to highlight the failure prominently in the main stream. Because the single chat log serves as an orchestration hub tracking numerous ongoing processes, critical notifications get lost within a sprawling, messy wall of text.

```mermaid flowchart LR accTitle: ChatGPT Feature Sprawl accDescr: Diagram mapping the fragmented feature surface area inside modern ChatGPT that contributes to UI clutter. subgraph UI ["ChatGPT Workspace"] Chats["Standard Chats"] Codex["Codex Interface"] Spaces["Spaces (Pages & Sites)"] Scheduled["Scheduled Tasks"] DotsAgent["dots Always-On Agent"] end User["User"] --> UI ```

This reflects a broader design crisis emerging within ChatGPT: - **Feature clutter:** The app has accumulated disparate modules including standard chats, Codex, Spaces (encompassing pages, sites, and image generation), scheduled automated tasks, and voice interactions. - **Cognitive overhead:** A once minimalist interface is now overwhelmed with competing toggles, modes, and distinct execution environments. - **Voice as a workaround:** Communicating directly with dots (named "Zoe") via voice mode often proves cleaner than navigating the crowded screen, but the interface risks alienating non-technical users if product teams do not prioritize aggressive simplification.

---

[16:17](https://www.youtube.com/watch?v=z5X1eMU6isI&t=977s) Competitive Landscape and the Future of AI Collaboration

OpenAI faces stiff competition from alternate agent ecosystems like Grok Bot, which recently debuted multiplayer "team bots." While OpenAI maintains high-quality execution harnesses, it remains primarily a single-player interface where an individual manages private agent threads.

The ultimate competitive battleground between OpenAI, xAI (Grok Bot), and Anthropic lies in building an AI-native alternative to Slack. The winning platform will seamlessly blend human team members and autonomous AI agents in shared workspaces without inheriting the noisy, overwhelming notification fatigue of legacy workplace chat apps.

- Dots is currently available for ChatGPT Pro subscribers. - While OpenAI continues rapid iteration, managing complex multi-agent workflows remains a work in progress balancing UI density against autonomous agent utility.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=z5X1eMU6isI&t=0s) dots is OpenAI's new always-on agent that lives inside ChatGPT. — Peter Yang

> [03:18](https://www.youtube.com/watch?v=z5X1eMU6isI&t=198s) sell the finished work — Zoe

> [03:18](https://www.youtube.com/watch?v=z5X1eMU6isI&t=198s) A-minus — Zoe

> [03:18](https://www.youtube.com/watch?v=z5X1eMU6isI&t=198s) Don't disappoint me — Peter Yang

> [07:38](https://www.youtube.com/watch?v=z5X1eMU6isI&t=458s) the part after an impressive demo — Peter Yang

> [09:28](https://www.youtube.com/watch?v=z5X1eMU6isI&t=568s) Let's try to do some ambitious things. What else can you do? Give me your top five. — Peter Yang

> [11:26](https://www.youtube.com/watch?v=z5X1eMU6isI&t=686s) PM Workbench — Zoe

> [11:26](https://www.youtube.com/watch?v=z5X1eMU6isI&t=686s) borderline slop — Peter Yang

> [16:17](https://www.youtube.com/watch?v=z5X1eMU6isI&t=977s) team bots — Peter Yang

## Mentioned Peter Yang, OpenAI, ChatGPT, dots, Zoe, Grok Bot, X Scout, Muse, xAI, Anthropic, Claude 3.5 Sonnet, Slack, X, Maven, Colin, Pokémon Red, DevDay, YouTube, App Store, Riverside, Google, Substack, Mickey, Mac

## Video description Copy my best AI workflows to save time and automate busywork: https://www.behindthecraft.com 🤝 Subscribe to my practical AI newsletter read by 150K+ professionals: https://creatoreconomy.so

dots is OpenAI's new always-on agent that lives inside ChatGPT. I tested it for a few days to run my one-person business, and it flagged a five-figure sponsorship issue I missed, got my fitness app ready for launch, and built a landing page for a paid product. In this video, I compare dots to Grok Bot and Muse and share my honest take on where it's useful and where it still falls short.

What I cover: (00:00) What dots is and what I tested (00:42) How dots works inside ChatGPT (01:34) dots vs. Grok Bot and Muse (03:18) Research on X and in Slack, plus memes (07:38) Analyze 1,800 YouTube comments (08:29) Launch my fitness app and audit sponsors (11:26) Build a paid product and grow my newsletter (13:29) Where dots falls short and my verdict

Where to find me: X: https://x.com/petergyang LinkedIn: https://www.linkedin.com/in/petergyang/

Subscribe to this channel - more tutorials coming soon!

[Source](https://www.youtube.com/watch?v=z5X1eMU6isI)

中文简报

OpenAI Dots 揪出 1 万美元账单错误，但受困于 ChatGPT 界面问题 OpenAI 的 Dots 在作为“幕僚长”管理后台工作线程、审计业务记录和发布代码方面表现出色，但严重的 ChatGPT 界面杂乱和静默工具故障使其尚无法取代人工监督。 OpenAI 的 dots 作为一个全天候运行的协调器，可跨本地和云端环境运行，并原生调用 ChatGPT 现有的记忆和工具功能。在实际的个人业务测试中，它取得了重大成果：标记了一张未开具的发票和一笔 1 万美元的赞助账单差异、修复了应用代码漏洞，并汇总了 1,800 条受众评论。然而，它在生成需要微妙表达的产品文案时表现不佳——产出了千篇一律的“劣质内容（slop）”——并且在链式工具崩溃时出现了静默任务失败。Yang 总结道，虽然 dots 作为跨线程委派子任务的“幕僚长”表现优异，但 OpenAI 必须解决日益严重的界面臃肿问题，并向多人协作 AI 工作区转型。 OpenAI 面向 ChatGPT Pro 订阅用户推出了全天候智能体 dots，可通过桌面集成在 24/7 云端环境和本地 Mac 客户端之间运行。 (0:00) 在针对个人创作者业务的测试中，dots 进行了一次赞助审计，找出了五个关键的收入问题，包括一张未开具的 Riverside 发票，以及与 Google 之间一笔未解决的 1 万美元账单差异。 (9:28) Dots 最适合作为高级协调器或“幕僚长”，将专门的任务委派给次级 ChatGPT 线程，而不是在单个对话中直接执行所有操作。 (9:28) 尽管具备强大的协调和分析能力，dots 仍会出现缺乏明确报错提示的任务流水线故障，并凸显了 ChatGPT 平台内部日益严重的界面臃肿问题。 (13:29)

Reading summary

According to Peter Yang's hands-on evaluation, OpenAI's "dots" agent functions effectively as an executive Chief of Staff within ChatGPT by coordinating sub-threads to resolve operational oversights and synthesize business data. However, its practical utility is constrained by brittle multi-step integrations, mediocre long-form copy generation, and an increasingly cluttered interface. Dual Execution and Sub-Thread Orchestration (0:42) Dots functions as an always-on coordinator integrated directly into ChatGPT rather than a single-prompt execution tool. Operating simultaneously across a local Mac laptop and a continuous 24/7 cloud environment, the agent—tested under the persona "Zoe"—relies on native ChatGPT memory, voice calling, and a Slack integration. Rather than executing complex tasks directly, dots coordinates dedicated sub-threads, acting like a project manager supervising subordinate worker threads. This delegation model proved effective in technical preparation. When assigned to prepare a fitness application for release, dots spawned one sub-thread to prompt image generators for App Store promotional layouts and routed diagnostic prompts into the core build thread to fix bugs in the codebase. By running live tests of the application, the agent captured interface screenshots, layered promotional headlines over them, and prepared a full visual suite for store submission. Measurable Financial and Diagnostic Business Impact (9:28) The agent demonstrated immediate utility across financial auditing and audience research. Because dots maintains access to historical conversations without requiring re-authentication, it autonomously audited past sponsorship records. The review discovered that sponsor Riverside had not been billed for completed spots and flagged an unresolved $10,000 billing mismatch on Google sponsorship campaigns, effectively acting as an automated revenue manager. In data synthesis tasks, dots parsed feedback across more than 300 YouTube videos, taking roughly 30 minutes to synthesize over 1,800 comments. It concluded that channel growth required disclosing operational API costs, designing tutorials runnable from blank setups, and demonstrating production maintenance beyond initial demos. Similarly, within a Slack community, the agent scanned discussion archives in seven minutes, assigned an "A-minus" grade, and outlined engagement initiatives alongside messaging revisions for an educational course. Content Limitations and Brittle External Integrations (11:26) Despite strong analytical capabilities, dots struggled with autonomous product creation and multi-step tool execution. When directed to build a paid $99 digital product concept called "PM Workbench" featuring four sample artifacts, the agent generated polished visual assets, but Yang evaluated the product copy as "borderline slop" that required substantial human revision. Operational friction also appeared during integrated workflows. An attempt to convert a Riverside interview transcript into a Substack newsletter draft stalled due to tool-availability and authentication errors in Substack. Crucially, dots failed to highlight these failures clearly in the main stream, burying error notifications inside a sprawling text log. Interface Clutter and the Competitive Field (13:29) The agent's text-heavy reporting reflects an emerging interface crisis inside ChatGPT. As the application accumulates disparate features—including standard chats, Codex, Spaces, scheduled automations, and voice—managing dense agent logs creates substantial cognitive overhead. While voice mode offers a cleaner conversational alternative, Yang notes that managing complex workflows risks alienating non-technical users unless OpenAI simplifies the layout. Competitively, dots targets knowledge work rather than consumer logistics like Muse, which focuses on outbound telephony and marketplace bookings. While Grok Bot provides multi-agent team swarms, dots currently operates as a single-agent setup whose primary advantage is zero-setup access to ChatGPT's accumulated context and authenticated tools. Yang argues that the broader competition between OpenAI, xAI, and Anthropic will ultimately center on building an AI-native alternative to Slack that unifies agents and human collaborators without notification clutter. Currently available to ChatGPT Pro subscribers, dots demonstrates that autonomous agents can uncover missed revenue and coordinate complex tasks, but reliable end-to-end autonomy will require OpenAI to reduce interface sprawl, improve failure visibility, and expand beyond single-agent architectures.

Source quotes

[0:00] Peter Yang: dots is OpenAI's new always-on agent that lives inside ChatGPT. [3:18] Zoe: sell the finished work [3:18] Zoe: A-minus [3:18] Peter Yang: Don't disappoint me [7:38] Peter Yang: the part after an impressive demo [9:28] Peter Yang: Let's try to do some ambitious things. What else can you do? Give me your top five. [11:26] Zoe: PM Workbench [11:26] Peter Yang: borderline slop [16:17] Peter Yang: team bots

Original transcript

[0:00] This is Dots Open AI's new always-on [0:03] agent for ambitious work. I've been [0:05] testing it for the past few days to run [0:06] my business and it flagged a five-figure [0:09] sponsorship opportunity that I missed, [0:11] found a bunch of bugs in my fitness app, [0:13] and also got it ready for launch with [0:15] App Store screenshots, suggested a paid [0:17] product that I should offer, and even [0:19] built a landing page, and also posted [0:21] some dank memes on Slack. Now, I'm sure [0:24] you're wondering, is Dots better than [0:26] Grokbot and Muse? [0:28] In this video, I'm going to answer that [0:29] question and also walk through 10 [0:31] different things that I tried to do with [0:33] my Dots and both the good and bad [0:35] results. But first, let me start by [0:37] showing you how Dots works and how it [0:39] compares to the other personal [0:41] assistants. Okay, so this is the new [0:43] ChatGPT app and you can see right at the [0:45] top there is a new pinned thread for [0:48] your Dot. Now, I named my Dot Zoe and [0:52] just like Grokbot, I can customize the [0:54] name and also the appearance of my Dot. [0:57] So, for example, I've given it a [0:58] triangle appearance with some apples. [1:01] All right. Now, down here on Dot's [1:03] profile, you can see that there's a few [1:05] buttons, right? I can call my Dot using [1:07] ChatGPT voice, I can uh connect with her [1:10] through Slack, and also potentially send [1:12] her email, although this feature is [1:14] currently unavailable. [1:16] And Dot is also able to connect to two [1:18] different computers. It has a remote [1:21] cloud computer that runs 24/7 [1:22] [clears throat] [1:23] and it's also connected to my laptop [1:25] along with ChatGPT. And here we have an [1:28] activity menu, which is all the [1:29] different tasks that I got my Dot to [1:31] work on that I'm going to walk through [1:33] later in this video. Now, let's talk [1:35] about how this product compares against [1:37] Muse and Grokbot. In my opinion, it is [1:40] not competitive with Muse at all. Muse [1:42] is focused on consumers, [1:45] like people to call you as businesses, [1:46] book plane tickets for you, sell stuff [1:48] on Facebook Marketplace, while Dots is [1:51] very much designed for work use cases. [1:53] So, it's directly competitive with [1:55] Grokbot. [1:56] So, those use cases include, you know, [1:57] doing competitive research, helping you [1:59] understand your business, and more. [2:01] Now, each of these different apps have [2:04] different killer features. So, for me [2:07] use, for example, in my opinion, the [2:09] killer feature is being able to actually [2:11] call US businesses, right? So, recently, [2:14] for example, I used me use to order a [2:16] birthday cake for my wife. And for [2:18] Grokbot, I think the killer use case is [2:21] just having a team of bots here on the [2:23] left and be able to get one bot to tell [2:26] the other bots to do work, and also add [2:28] multiple bots to the same chat threads [2:30] to kind of work off each other. Now, [2:32] with dots, you can only have one dot to [2:34] start, although I believe Open AI says [2:37] more dots are coming soon. [2:39] And [2:40] so, really, I think the killer use case [2:43] here for dots is just native chat GPT [2:46] integration, right? Basically, my Zoe [2:49] dot already has all the relevant memory, [2:52] conversation history, connected apps, [2:55] and everything else that I've set up [2:56] already with chat GPT. So, you can see [2:58] here that the onboarding experience is [3:00] very straightforward. I don't have to [3:01] reconnect my email or do other things. [3:04] Everything's connected by default, and [3:07] it's basically set up for you right off [3:08] the bat. So, that's the benefit of dot [3:11] being natively integrated into chat GPT, [3:14] especially if you use chat GPT a lot [3:15] already, just like me. All right, so, [3:17] the first use case that I tried with my [3:19] dot is scout trending practical AI posts [3:23] on X for me, uh deliver in a chat. [3:27] Basically, the same prompt that I had my [3:28] Grokbot do, right? And dot basically, uh [3:31] I think used the sign browser on my [3:33] back, my local Mac, to browse and find [3:37] all these different tweets that I think [3:38] are interesting. And let's switch over [3:40] to Grokbot, and I have a dedicated [3:42] Grokbot to do this. It's called X Scout, [3:45] and basically, it also uses browser use [3:47] to find the top tweets. Now, these two [3:50] different bots found two completely [3:51] different sets of tweets. [3:53] So, honestly, I I can't really tell [3:55] which one is better or not. But, [3:56] basically, you can use both to research [3:59] tweets for you. Okay, so that was the X [4:01] research use case. Now, let me actually [4:03] jump over to some more interesting use [4:04] cases, right? And for this, I'm going to [4:06] go over to Slack. [4:08] So, to connect Slack to your dot is very [4:10] simple. Just click this button and sign [4:12] in and it'll be connected. All right, so [4:14] I have a Slack community for my [4:16] newsletter and here's a direct thread [4:18] with Colin, who is a AI instructor who I [4:21] really respect and a good friend. [4:23] Colin has an AI course that you see here [4:26] on Maven. [4:28] So, what I asked my dot to do after [4:30] adding it to a DM thread is [4:32] "Can you do some deep research on Maven [4:34] courses similar to Colin's course and [4:36] tell us the top two or three things to [4:38] update Colin's landing page and make it [4:40] a no-brainer to take?" All right? [4:43] So, you can see here it started doing [4:44] research and just, you know, 6 minutes [4:47] later, it has the results. [4:51] So, it told Colin to basically sell the [4:53] finished work. Here's the proposal [4:54] headline. [4:56] And it did a bunch of research on other [4:57] similar courses. And here's the top [5:00] three things that Colin should do. So, [5:01] you can use dot in Slack to do [5:03] competitor research. The next thing I [5:05] asked my dot, Zoe, to do is to look back [5:08] on all threads in this community and [5:11] give me your letter grade of how good I [5:13] am at running this community. Be nice [5:15] and also give me top three things to [5:16] improve the community. All right. [5:18] >> [gasps] [5:18] >> So, Colin thinks Zoe's a little bit [5:19] slow, but I don't think so. I think, [5:21] basically, in 7 minutes, it actually [5:23] went through all the previous Slack [5:25] threads in this community. [5:27] And [snorts] uh it actually gave me a A- [5:29] minus, which is a better grade than I [5:31] would give myself. [5:32] But, uh three ways to get an A: Make [5:34] meeting other members part of the [5:35] product. So, run a monthly session with [5:38] 5 to 10-minute member demos [5:40] followed by small group conversations. [5:43] Make the best community knowledge easy [5:44] to find again. [5:46] Build on the portal link durable [5:47] lessons. [5:48] And give it and give every help request [5:50] a next step. Okay, so again, dots and [5:53] ChatGPT is very good just like analyzing [5:56] a lot of information and giving you [5:58] actionable takeaways, right? [6:01] But let me actually move on to the third [6:03] use case, which is probably the most fun [6:05] use case. And here I asked my dot to [6:08] generate five amazing memes showing [6:10] Peter and Colin doing various activities [6:12] based on this community and what you [6:13] know about these two people. [6:15] Don't disappoint me, okay? That's what I [6:17] said. [6:18] And here's a bunch of memes that Zoe [6:20] came up with. [6:21] So, come for the AI skills, [6:23] stay because Peter locked the meme [6:25] gates. [6:26] We saved the community history, but more [6:27] importantly the memes. [6:29] And here's uh I really enjoyed this one, [6:32] which is I'm always trying the latest AI [6:34] tools, and Colin is much more uh [6:37] principled. He, [6:39] you know, is not always chasing [6:40] something new. There's some memes here. [6:42] And then I asked uh Zoe to give me more [6:45] memes, this time looking at recent AI [6:48] news. [6:49] And [snorts] here doesn't seem as good [6:51] as the personal memes they made for me [6:52] and Colin. This one is about what? [6:55] Claude Sonnet 5.5 beating Pokémon Red. [6:58] These memes just aren't as funny as the [7:00] previous personal ones. [7:02] And then I asked it to speculate and [7:04] predict what launch at DevDay. So, this [7:06] video is coming out on DevDay. [7:08] But uh here's some predictions that Zoe [7:10] made. [7:11] Yeah, the these also aren't as good [7:13] other than what it did with my face. [7:15] All right? But like, even talking to [7:17] OpenAI, they actually have a dedicated [7:18] Slack channel internally where they have [7:21] all of their dot bots for all their [7:24] employees inside a channel and just [7:26] constantly sharing and replying to each [7:28] other's memes. And they told me it's one [7:30] of the most interesting channels that [7:31] they have internally. So, again, if you [7:34] have this stuff at work, try creating a [7:35] meme channel for your dots. All right, [7:38] let's go back to some more serious [7:39] topics. So, back in the thread, I asked [7:43] my bot to go read every comment I've [7:44] ever received on YouTube over the last 6 [7:47] months [7:48] and give me the top five actionable [7:50] themes to grow my channel faster, right? [7:52] So, to do this, it had to look up over [7:54] 300 videos of comments. So, it took [7:57] actually about 30 minutes to come up [7:59] with this. You know, if I scroll down [8:01] here, here is what it came back with. [8:03] So, my take, your clearest growth [8:05] opportunity is making impressive demos [8:06] easier to produce and trust. I [8:08] prioritize these five changes. [8:10] Show how much the workflow actually [8:11] costs. [8:13] Make every tutorial work from a blank [8:14] setup. Show the part after an impressive [8:16] demo, and so on and so forth, right? So, [8:19] it went through over 1,800 comments to [8:21] come with these five recommendations. [8:23] So, that's pretty useful again to kind [8:25] of summarize information and get [8:26] actionable takeaways. [8:28] All right, next I asked it to take a [8:31] look at my fitness app, build thread. [8:33] So, down here I have a build thread for [8:34] my fitness app. Understand the code base [8:37] and create six App Store screenshots [8:40] to get this app launch-ready. So, here, [8:42] instead of doing the work in its thread, [8:44] it's actually passing instructions over [8:46] to this build thread to do the work. So, [8:48] let's take a look. So, you see here that [8:49] it actually started a new thread to get [8:52] ChatGPT images to generate App Store [8:54] screenshots. [8:55] And here, it's asking the build thread [8:57] to look for any bugs or issues, and then [9:00] it fixed all these issues by itself. [9:02] Now, let's go back to our Zoe thread. [9:04] Here are the App Store screenshots that [9:06] it generated, right? I think this looks [9:08] pretty pretty good. I guess you go one [9:10] at a time. These are the actual [9:11] screenshots, by the way, of my app. So, [9:14] it must have like run a test of this app [9:16] and takes taken screenshots to generate [9:17] this, and then added its own headers [9:19] here, too. [9:21] So, the next step I can do is just tell [9:23] her to get it ready for the App Store [9:24] launch. Okay, so the reason I want to [9:26] show this is because Dot is very much [9:28] meant to be an orchestrator's thread. [9:31] You shouldn't really get Dot itself to [9:32] do work. You should get it to [9:34] orchestrate the other ChatGPT threads [9:36] that you have to get work done and [9:38] report back on their progress to dot. [9:41] So, in some ways it's kind of like a [9:42] chief of staff. It's kind of like a [9:43] manager of your other threads. [9:46] And one of the best prompts that you can [9:48] give your dot to kind of get it to [9:50] orchestrate these other threads is let's [9:53] try to do some ambitious things. What [9:55] else can you do? Give me your top five. [9:57] Because dot is part of chat GPT and it [10:00] has memory of all your other [10:01] conversations, it can actually suggest [10:03] some really, really relevant things, [10:05] right? So, the first thing is like, hey, [10:07] build a new pay product for your [10:08] audience's problems. Take a video from [10:10] idea to publish ready. Run a sponsorship [10:13] revenue audit. Take the app through a [10:15] full launch rehearsal and turn your [10:17] community into a hands-on experience. [10:21] And so, here I replied, let's go ahead [10:23] and do all these things. And probably [10:25] the most relevant thing that my dot has [10:28] done for me in this past couple days is [10:30] actually number three, running a [10:32] sponsorship revenue audit. So, you know, [10:35] I run a one-person business and I'm not [10:36] always on top of what's going on with my [10:38] sponsorships. And I really should be [10:40] because it is my main revenue driver. It [10:42] is what pays the mortgage and the bills. [10:44] And right off the bat, dot found five [10:46] things wrong with my sponsorships. So, [10:48] number one, for Riverside, which is a [10:50] sponsor, it seems that I forgot to bill [10:52] them. So, I should send them an invoice [10:55] uh to not lose any revenue. [10:57] And then there's some confusion about [10:59] Google sponsorships including a $10,000 [11:01] discrepancy uh between billing that I [11:03] should resolve, right? So, just having [11:06] it kind of like proactively search [11:08] through everything and bring the stuff [11:10] to me while I may not be aware of it is [11:13] extremely useful. And you can imagine [11:14] that I can tell my uh dot bot to [11:16] basically keep track of my sponsorships, [11:19] kind of be a sponsorship manager, [11:21] and kind of hold me accountable and help [11:22] me keep track to not lose out on revenue [11:25] opportunities. All right, so let's [11:26] scroll back up here. And another thing [11:29] that dot started working on for me is to [11:31] build a new paid product based on my [11:34] audience's problems. And it proactively [11:35] suggested this. So, I said, "Hey, okay, [11:37] let's go ahead and build it. And I'm [11:38] really curious to see what it is." [11:41] Okay, and if you scroll down here, uh it [11:44] says the website is ready for review. It [11:46] made a PM workbench a $99 workflow kit [11:49] concept with four interactive examples. [11:52] And it made this as a ChatGPT site. So, [11:55] I took a look at it and I told it, "Hey, [11:56] this thing is hella boring. Make it much [11:58] better. Use ChatGPT images just to add [12:00] some engaging images and so on and so [12:02] forth, right? So, let's take a look at [12:03] the actual website that it generated [12:05] after this feedback. So, this is the [12:07] ChatGPT site generated. It is a digital [12:09] field kit for product managers. [12:11] All right, so it has customer evidence, [12:13] competitor intel, and so on and so [12:15] forth. Open a real example. [12:18] So, the real example is there's an [12:20] evidence brief, it has a competitive [12:22] scan, [12:23] stakeholder update, and PRD red team. [12:26] I think this website looks pretty good. [12:28] I I I I love the ChatGPT images. But the [12:31] copy, as someone who has done PM for 10 [12:33] years, doesn't seem quite right. You [12:35] know, it it seems kind of borderline [12:36] slobbish. So, I don't think [12:40] at least in this example, my dot is able [12:42] to one-shot generate a [12:44] revenue-generating product. I think I'm [12:45] going to give it a ton of feedback on [12:46] this to see if this is even a good idea [12:48] or not. But going back to our thread, [12:51] you know, I love how it's proactive and [12:53] it even suggested this idea. I also [12:56] asked dot to analyze my Substack [12:58] newsletter numbers and history and give [13:01] me some actionable suggestions to grow [13:03] it faster. And it gave me, you know, [13:05] again, these five top five moves. Fix [13:07] the membership buying journey, double [13:09] down on newsletter recommendations, and [13:11] so on and so forth. And this is based on [13:12] data that it pulled using a browser on [13:14] his computer just browsing around my [13:16] Substack stats dashboards and recording [13:18] the numbers. So, yeah, dot and AI, [13:21] obviously, is very good at just [13:22] synthesizing a lot of research, doing a [13:24] lot of manual work, navigating pages, [13:26] and then pulling it together an [13:27] actionable report for you. Okay, so I [13:29] also want to be transparent and show you [13:31] some examples where Dot struggled. So, [13:34] one example is I asked it to take the [13:37] Riverside transcript of my interview [13:39] with Mickey, and then use my skills like [13:41] podcast prep, podcast edited to [13:44] basically draft a newsletter post. [13:46] And here it kind of struggled is that [13:47] the blocker is tool availability. Dot [13:50] verify Substack authentication. Blah [13:52] blah blah. [13:53] And uh [13:54] the task kind of failed. Okay, and back [13:56] in Dot, as far as I can tell, it didn't [13:58] actually tell me that the task failed. [14:01] Uh it might have told me somewhere in [14:02] this long thread, but you know, there's [14:04] so many different work streams going on [14:06] here in this thread that it's very hard [14:08] to find what's going on. You know, I've [14:09] already gotten used to If If you look [14:12] carefully here, I already have a chief [14:13] of staff thread here that I'm using to [14:16] orchestrate all my other threads. I [14:18] think Dot Chat GPT's official chief of [14:21] staff thread, right? So, just from this [14:23] conversation alone, [14:25] you can see how messy it is to actually [14:26] look through it. It's really meant for [14:28] you to spin up five different threads or [14:31] get five instances threads to do work. [14:33] And you know, use voice or look at this [14:35] chat to keep track of how things are [14:36] going, to have Zoe or your Dot report [14:38] back, so you can give more commands. [14:41] Think of it as you are the manager of [14:42] your Dot, and your Dot manages basically [14:45] everything else in Chat GPT. And this is [14:47] actually the thing that I'm most [14:48] concerned about with Chat GPT. [14:51] I've been using it as my default harness [14:53] for the past 4 or 5 months. [14:56] And overall, it's incredible. The voice [14:58] mode, computer use, browser use, all [15:00] works really really great. [15:02] But you can see here that the UI is just [15:04] getting really really complicated. There [15:05] is Chat GPT, there is Codex that you can [15:07] switch back and forth. There is this new [15:09] space [15:10] experience where you have pages, sites, [15:14] and images, and other things you can [15:15] navigate. [15:17] And there's other things like schedule [15:18] tasks and more. And in my opinion, it's [15:21] just getting really, really overwhelming [15:23] how many different threads and buttons [15:25] and toggles there are in ChatGPT [15:28] as opposed to like the simple experience [15:29] that we had before. So, maybe the [15:31] solution here is actually just using [15:33] your. dot [15:34] to could basically orchestrate [15:35] everything else. [15:36] You know, maybe I don't have to look at [15:37] this UI at all and I can just have a [15:39] call with Zoe to get work done, which is [15:41] often what I do. But, I think OpenAI [15:44] should be very careful. You know, [15:45] obviously with AI, you can ship all [15:48] kinds of great features into this app. [15:50] But, what started as a simple streamline [15:52] experience is getting increasingly more [15:55] and more complicated to use. I think the [15:56] designers and the PMs and the engineers [15:59] kind of should hold their ground on [16:00] making this as simple and easy to use as [16:02] possible. [16:03] And and I know OpenAI is aware of this [16:06] and they're working on simplifying this. [16:07] But, if someone like me who's just [16:09] talking about AI and all these new AI [16:11] tools all the time is getting confused, [16:13] the average ChatGPT user is going to be [16:15] incredibly confused. Okay, so I hope you [16:17] found this tour of dots useful. You [16:20] know, I know the OpenAI team worked [16:22] incredibly hard to get dots to this [16:25] moment. But, right now they're still [16:27] playing a feature catch-up to Grok Bot. [16:28] So, for example, Grok Bot recently [16:30] launched team bots that I'm sure OpenAI [16:32] is also working on. And I also think [16:34] OpenAI should think about, "Hey, should [16:36] we make ChatGPT kind of a multiplayer [16:39] agent-first Slack experience [16:42] or should we still do stuff in Slack?" [16:44] But, I think the reason that OpenAI, [16:46] Grok Bot, and I'm sure Anthropic at some [16:48] point is all competing here for work use [16:51] cases is there's just like a really [16:53] large opportunity here to build an [16:56] AI-native version of Slack where humans [16:59] and agents can collaborate together. Cuz [17:02] right now, like ChatGPT is still [17:03] primarily a single-player experience [17:06] with me and my bots. [17:08] But, you know, if teams and bots can all [17:10] collaborate each other [17:11] without being as overwhelming as Slack [17:13] currently is, that can be a huge [17:16] differentiator. So, as you can see from [17:18] the rest of that day announcements, [17:20] OpenAI is clearly shipping. It's super [17:22] exciting, but also they should be really [17:24] careful. I think ChatGPT and Codex again [17:27] has some of the best interfaces in AI [17:30] and shipping this many products at once [17:33] is making this whole app a little bit [17:35] more confusing to use. [17:37] But, I believe in OpenAI. I'm sure [17:38] they're going to simplify this soon. [17:40] And Docs is available for ChatGPT Pro [17:43] subscribers. So, give it a try and let [17:46] me know in the comments what you think. [17:48] And as always, please like and subscribe [17:50] if you enjoy more practical and no hype [17:53] tutorials like this.