How an Agent Should Stop Itself Acting Without You
An agent that can message people, move calendar events and file disputes for you is only useful if it can't do any of that behind your back. Here are the design rules we think every agent app should be held to — and the questions to ask before you grant one a single permission.
A safe AI agent shows you the full plan before anything runs, keeps each permission separate and off by default, never acts in the background beyond what you approved, pauses on money and commitments, and keeps a record you can read.
Why agents are different from other apps
Most apps do exactly what you tap. An agent decides what to tap. That shift — from you choosing every step to software choosing steps toward a goal — is what makes it powerful, and it's also where the risk lives. The failure you should worry about isn't a wrong answer. It's a right-sounding action you never approved.
Rule 1: confirm the plan, not just the first step
For a one-off action, "are you sure?" is enough. For a multi-step job — rebooking a flight, messaging a hotel, telling a driver you'll land late — approving each step separately is tedious, and approving only the first step is dangerous. A good agent shows the whole plan in plain language, once, and nothing runs until you say yes. If the plan changes mid-way, it comes back and asks again.
Rule 2: permissions per capability, off by default
Messaging, calendar, phone calls and device settings are different levels of trust. They should be switched on individually, they should start switched off, and the agent should only ever propose actions that match what you've granted. If you haven't given it calendar access, it shouldn't be able to quietly "just check" your calendar.
Rule 3: nothing in the background without your say-so
A job like "text me when the contractor replies, then add the quote to my budget sheet" involves waiting. That's fine — as long as you approved that exact plan up front, and the agent doesn't quietly widen it. Waiting on a condition is not permission to do something new when it fires.
Rule 4: pause on money, deadlines and commitments
Some actions are cheap to undo and some aren't. Sending a reminder is cheap. Accepting a new contract, paying an invoice or telling a landlord you're leaving is not. A careful agent treats those differently: it prepares everything, then holds for an explicit yes on the specific step that commits you.
Rule 5: a record you can actually read
You should be able to see what you asked for, what the agent understood, what it did and when. A readable history is how you catch a misunderstanding early, and it's how you build the kind of trust that lets you delegate bigger jobs over time.
Five questions to ask before you grant an agent a permission
- Does it show me the full plan before anything runs?
- Can I turn each capability on and off separately?
- Does it ever act when I'm not looking?
- What does it do with money, deadlines and commitments?
- Can I read back what it did, and for how long is that kept?
How StareBrain approaches this
StareBrain is built around these rules: you see the plan before it runs (texts always ask, and auto-confirm is optional and off by default), each capability can be switched off, the agent pauses before irreversible taps, and your command history stays on your phone and clears after 7 days. It hasn't launched yet, so none of this is a claim about results — it's the standard we're building to, and the one we'd suggest you hold any agent app to.
This post describes design principles, not measured results. See our privacy policy for how data is handled.
Questions
- Is it safe to let an AI agent act on my phone?
- It can be, if the agent shows the full plan before anything runs, keeps each permission separate and revocable, holds for your approval on money and commitments, and keeps a readable record of what it did.
- What should an AI agent never do without asking?
- Anything that commits you: paying, accepting a contract, sending a firm or irreversible message, or widening a task you already approved. A careful agent prepares these but waits for an explicit yes.
- Can I revoke an agent's permissions later?
- You should be able to. A well-designed agent lets you turn each capability off individually at any time, and loses that ability immediately.