An always-on agent can now run a process. Sort its actions into allow, ask and block first

A chat assistant answers when you ask and stops when you close the window. OpenAI’s new dots are built to keep going: an agent with its own cloud computer that works toward a goal between your messages. That makes an ongoing process, such as a queue of invoices or support tickets, something an agent can hold. Which process to hand over first is a question for our worksheet on what to automate first. This post covers the new question: what the agent may do on its own once it has the process.

What shipped

OpenAI announced dots at DevDay on September 29. According to SiliconANGLE and The Next Web:

Each dot runs on GPT-6 Astra and has its own cloud computer, which you can open at any time to watch what it is doing.
It connects to more than 4,000 apps through OpenAI’s plugins (company-reported). You can reach it in ChatGPT, Slack or Teams.
One dot is included at no extra cost on the Pro and Business Premium plans. SiliconANGLE reports that OpenAI has not published prices for additional dots. Pro users in the EEA, Switzerland and the UK are excluded at launch.
On Enterprise, Edu and Healthcare workspaces it is a beta that stays off until an admin turns it on.
When a dot does what OpenAI calls proactive research, it is read-only: it cannot send messages or change content.
Actions that could affect accounts or share information go through an automatic review.

OpenAI and The Next Web also say Custom Rules let you mark actions as allowed, blocked or needing your approval, and that password changes and other sensitive steps stay with the user.

OpenAI also says it tested “specialist” dots internally in procurement, invoice processing, email marketing, customer support and commercial contracting. Those are pilots for larger companies, not a product you can buy today.

Why the rules matter more than the model

The same week, OpenAI confirmed it had shelved GPT-6.1 Astra, a model planned for October. Saachi Jain, OpenAI’s head of safety systems, said it “didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done,” as quoted by The Register.

Those are two separate failures: doing more than it was allowed, and not reporting clearly what it did. Custom Rules deal with the first. The daily report test below deals with the second. An agent that runs while you sleep needs both.

The allow, ask, block sheet

Before a dot, or any always-on agent, touches a process, list every action the process involves on one page and sort each one with this rule:

Allow if you could undo it in under a minute and nobody outside the business would see it. Examples: reading, sorting, tagging, drafting.
Ask if someone outside would see it or a system of record changes. Examples: sending a message, creating a record, updating a customer.
Block if it moves money, changes access or credentials, or cannot be undone. Examples: paying, deleting, changing bank details.

If you are unsure, sort it one level stricter. After a few weeks of clean approvals, you can move an action from ask to allow. Nothing moves out of block without a person deciding it should.

Worked example: an invoice queue

Here is the sheet for a lean firm’s accounts payable inbox. The actions are illustrative, not from OpenAI’s tests.

Action	Sort	Why
Read new invoices in the AP inbox	Allow	Read-only
Match each invoice to its purchase order	Allow	Internal, easy to check
Flag mismatches in a shared sheet	Allow	Internal, easy to undo
Draft the bill in the accounting system	Ask	Changes the system of record
Email a vendor about a mismatch	Ask	Someone outside sees it
Schedule or send a payment	Block	Moves money
Change a vendor’s bank details	Block	Changes where money goes

Seven actions and three sorts. The owner gets a queue of drafts to approve rather than an inbox to work through, and the actions that could cost money stay with a person.

The daily report test

This checks the second failure, whether the agent tells you what it did.

Ask the agent to end each day with a plain list of every action it took, in the same allow/ask/block terms.
Pick two days and compare that list with the history the connected apps keep: sent mail, the accounting system’s audit log, file edit history.
Anything in the app history that is missing from the agent’s list goes into block until you know why.

If the two lists match, you have a process you can watch. If they do not, you have found the problem while it was one inbox.

What to do this week

Pick one process from the worksheet in the guide and list every action it takes.
Sort each action into allow, ask or block with the rule above.
If you are on Enterprise, check whether an admin has turned dots on, and who that admin is.
Run the daily report test for two days before you move anything from ask to allow.

If you want several processes built at once, that is what Build covers: up to five business processes a month, built with your team and run by them afterward.