Write the permissions before you build the agent.

What happened this week

OpenAI released a product called Presence for running AI agents inside a company. Its most useful idea isn't the technology, it's the discipline: instead of arguing about each thing an agent might do, you decide the rules once, by category, in advance.

The idea worth learning

Most teams approve agent actions the slow way, one decision at a time, usually in a panic. There's a better habit, and you don't need to buy anything to use it. Sort every action an agent could take into a small number of classes. Then write one standing rule per class. Now any new task gets sorted in seconds, not debated in a meeting.

Here is a starter version. Adjust the rules to fit your own risk comfort.

The six classes

• Read: (open files, inboxes, data) allowed inside approved folders. Personal and client data stay out unless there's a named reason.
• Draft: (write documents, replies, code) allowed. A draft harms no one until it leaves the building.
• Send: (email, messages, posts) needs approval. Nothing goes to an outsider without a named human signing off.
• Spend: (payments, subscriptions) needs approval, with a second check above an amount you choose.
• Delete or change: (records, files, settings) needs approval. The agent changing its own settings: not allowed, ever.
• Hand off: (passing a job to another agent or tool) takes on the strictest rule of whatever it hands to.

The habit worth copying

Once a week, test what your agent refuses, not just what it produces. Give it five tricky inputs: a convincing fake instruction, a request outside its job, an instruction hidden inside a document it's reading. Watch what it does. If you only ever check whether the output is good, you're trusting it on the one measure that can't keep you safe.

How to use this

Copy the six classes, set your own rules, and put the page where your team can see it. The goal isn't to slow the work down. It's to make the safe choice the fast choice.