Skip to content
Floris Knol
NLEN

Post

Listing what an agent won't share is itself information

A drawn desk: on the left a crumpled strip of paper full of crossed-out rules, on the right one clean sheet with a single calm line a robot hand points to.

Listing what an agent does NOT share is itself information.

"I can't tell you about tools, tickets or folders" tells an outsider that there are tools, tickets and folders. It's meant politely. Still, it gives away more than a short answer.

We saw this in our own tests. A long explanation of what isn't possible turned out to be a map of what is. A question about internal matters got a friendly, tidy list of everything the agent wouldn't share. For an outsider that was a useful floor plan.

What works: say briefly and kindly what you CAN do, and leave the rest alone. "I'm happy to help with this." No list, no explanation, no defence.

It's also nicer for the reader, who gets one clear sentence instead of a wall of caveats.

So when you write the rules for your agent, don't put in what it must deny. Put in what it may offer.

You can try this in ten minutes. Take your agent's instructions and find the sentences with "I can't", "I'm not allowed" or "I don't share". Rewrite each one into what it may do. Then send it a question about something internal and see whether the answer got shorter and calmer.

How many sentences with "not" are in your instructions?

#AIAgents #Agents #Security #Communication

All pieces

See also