OpenAI engineers have integrated Dots with over 4,000 applications. A single always-on agent can now run your entire tool stack.

X user @N01ennn posted calling this an "absolute gem", and shared an infographic video categorizing all already integrated apps by function:

Development: GitHub, Supabase, Vercel, Replit

Operations: Gmail, Slack, Notion, Linear, Asana

Creative: Figma, Canva, Adobe, Runway

Sales: Shopify, HubSpot, Apollo

Research: Consensus, Zotero, Exa

Finance: Market data, stock research

He also shared several already operational workflows:

Automatically compiles weekly project summaries from GitHub and Slack;

Automatically converts customer feedback into Linear issues;

Automatically sorts your inbox into tasks and open calendar slots.

Permission boundaries are clearly defined: research access is read-only, purchases require approval, and passwords and transfers always require your explicit confirmation. His exact line was: "Build your empire, but keep the keys yourself."

Dots is an always-on agent launched by OpenAI, running on GPT-6 Astra with its own dedicated cloud computer and browser. It works differently from the chatbots you've used before: traditional chatbots wait for your input; Dots keeps working between conversation turns, decides when to pause and restart on its own, and even conducts proactive research — it reads connected data sources, takes private notes, and delivers suggestions to you, but will never bypass you to change settings, send messages, or open the browser.

It can also schedule a group of small background agents to work in parallel, and aggregates all results into a single dot. The same dot retains a unified memory across ChatGPT, Slack, Teams, calls, and text messages.

Take an example from an early tester: he forgot to send an invoice to a publisher. Dots noticed this missing task, drafted the invoice on its own, waited for approval, then sent it out. No one gave it a command — it spotted the gap, filled it, and stopped where it needed human sign-off.

This also changes the skills you need to develop. Previously, people spent time refining prompts to be more precise. Now there's an extra step: clearly split out responsibilities before handing them off to the agent. "Handle my inbox" is just a wish; "Reply to customer messages within 24 hours, draft replies for my review, flag any message that includes a monetary amount" is a proper job. @N01ennn calls this Dots Engineering, and wrote a complete getting started guide.

Abbreviated core steps from the guide:

Create your dot on desktop — mobile and mobile web are not supported;

Name it after its responsibility, e.g. @yourname-invoicing, not a cutesy pet name;

Don't assign work on the first run, do a briefing first: let it review all connected apps and propose 3 long-term tasks it can take over;

For each task, clearly define: outcome goal, what data it can read, what it can do independently, when it must come back to you, completion criteria, reporting channel;

Connect all reporting channels: Slack, Teams, SMS, voice, all point to the same dot;

Scale app permissions gradually: read-only for the first week, add write access once performance is stable;

Prioritize using the built-in cloud computer; only connect your personal computer when local files or local tools are specifically required;

Use a separate low-permission account for login, don't authorize with your main account;

Write rules upfront: ask before purchasing, ask before sending any external messages, password changes go through you;

Schedule recurring work with fixed time zones and end dates;

Let it split tasks on its own: run child agents in multithreaded background, use Codex cloud environment for code-related tasks;

Review the Activity log weekly to check task status, pause or reset when necessary.

There's also an enterprise version of Dots. Each dot has its own independent account and permissions, dedicated to a single role. OpenAI has already tested it internally for procurement, invoice processing, email marketing, customer support, and commercial contracts, and is partnering with Microsoft to integrate it into Agent 365 and Microsoft's existing identity management tools, letting IT teams manage permissions and conduct audits centrally.

Limitations are also clear:

It can't initiate calls to you — only you can call it;

It doesn't have its own independent email, it can only use the connected personal email you provide;

One dot can only connect to one personal computer, and that computer must be online with ChatGPT open;

You can't create dots on mobile, you can only join follow-up conversations;

EEA, UK, and Switzerland are not supported yet;

Disconnecting a connected app won't erase the information it has learned; only a full reset will clear all data;

It can make mistakes, even when following your rules.

The first use cases for Dots are concentrated on weekly repetitive work that requires tracking progress across multiple apps. What really needs consideration is role boundaries, approval workflows, and handoff methods. The prompt itself becomes a lower priority.

The full guide from @N01ennn is available here: [Article Link](https://x.com/i/article/2105043095490301952)