OpenAI's Dots are always-on agents with their own computers. Their rulebook is the part worth copying

Always-on agents are here from OpenAI too. How they decide when to act alone is more interesting than the avatar.

What happened

At DevDay, OpenAI introduced Dots, always-on agents powered by GPT-6 Astra. Each Dot runs on its own cloud computer with a browser, so it can keep working in the background, per The Decoder. Users reach it through ChatGPT, Slack and Microsoft Teams, and it keeps context across those channels. Plugins connect a Dot to more than 4,000 apps.

When nobody is working with it, a Dot does "proactive research" using only read-only tools that cannot send messages, change content or control a browser. For active tasks, Custom Rules let users allow an action, require approval, or ban it. OpenAI says an automated check reviews any action that touches accounts or shares information against those rules. An early tester's Dot spotted a forgotten invoice, prepared it, and sent it after approval. OpenAI admits Dots can make mistakes and recommends double checking results with real consequences.

For businesses, OpenAI is piloting specialist Dots with their own identity and credentials, tested internally in purchasing, invoice processing, email marketing, customer support and contract management. TechCrunch reports OpenAI is working with Microsoft on Agent 365 security controls. Dots are available for Pro and Business Premium users in eligible markets.

My take

You do not need Dots to use its design. The pattern is solid for any agent I would put inside a real business:

1. Background mode is read only. Watching, summarizing and flagging are safe by default.
2. Every write action has a rule: allowed, needs approval, or banned. Written down before launch, not after an incident.
3. The agent has its own identity and credentials, so its actions show up in logs as the agent, not as your ops manager.
4. There is an activity view a human actually reads.

The invoice example is the right shape. The agent noticed, prepared, and waited. A human pressed send. That split is where most small teams should start.

What is the one recurring task you would hand to an agent like this first? Tell me in a message and I will sketch how I would scope it. romielwillautomate.dev