Gideon AwolesiShow what the agent is about to do
Before an agent acts, show people exactly what's about to happen: the action, who or what it affects, and anything that happens as a result. "Send this reply to 3 customers and mark their tickets as solved" is something people can check. "Processing your request" isn't.
For actions that are easy to reverse, the preview can be a short line people can glance at. For actions that aren't, like sending money or deleting something, make it a step people have to confirm.
Let people decide what it can do alone
Different people trust an agent with different things, and that trust grows over time. Let them choose, for each type of action, if the agent asks first or goes ahead. Start new users on asking first, and let them loosen it once they've seen it work.
Make the current setting visible where the agent acts, too. If people have to dig into settings to find out what the agent is allowed to do, they'll stop trusting it the first time it surprises them.
Make every action reversible, for long enough
Anything the agent changes should be reversible for long enough that people notice a mistake. For some actions, that's a few seconds before an email actually sends. For others, it's a history of changes people can roll back.
Where an action really can't be undone, say so in the preview, before it runs, not after.
Keep a record people can read
Give people a log of what the agent did, written the way they'd describe it themselves. When something goes wrong a week later, that log is how they find out what happened, and how your support team helps them.
Tell the people on the other end
When an agent sends an email or books something on someone's behalf, the people on the other end are interacting with AI too. Let them know in the message itself, and say who the agent is acting for. Since 2 August 2026, the AI Act has asked providers to let people know when they're interacting with an AI system, so this is part of the design, not an afterthought.