Before asking what an agent can do, ask what it may do

What is this agent allowed to do without asking? In almost every serious conversation about agents now, that question comes up before anyone asks what it can do. Buyers ask it in the first meeting and security teams ask it in the review. The people who will work next to the agent ask it on day one, usually as "can it send that without me seeing it?"

Is that just caution slowing things down? We don't think so. It is the market pricing risk correctly. An agent that acts is a different thing from a tool that suggests, and organisations have learnt, some of them the expensive way, that an agent with no limits is a risk nobody wants to insure.

What this means in practice is that permissions can't sit on a settings page any more. They are how the product is described, how it is built and what the user sees, all at once. The same sentence that closes the sale ("it drafts, you send") has to be enforced in what the tools are allowed to touch, and shown on the screen. When these don't agree, the customer finds out at the worst possible moment.

So that's where we start. Before the demo, before choosing a model and before the roadmap, we ask what this agent decides on its own, what it decides with a person, and what it never decides. The rest follows from the answers.

Working on this in your own product? See the four ways to start, or write to [email protected].