Oversee: keeping an AI agent in check
Limit what it can do, preview it, stop it, undo it.
Some AI agents don't just write text. They take real actions, like paying a supplier or changing prices. Oversee gives you control over one. A practice run shows what will happen before anything does, and you can stop the agent or undo its work at any point.
Live prototype: Oversee. Data is synthetic, and the company is fictional.
The problem: an agent's mistakes actually happen
When a chatbot gets it wrong, you get bad text. When an agent gets it wrong, something has already happened.
It pays the wrong supplier or emails the wrong list, and it's done before you read the summary. An "are you sure?" pop-up on every step doesn't help, because nobody reads fifty of them carefully. Most tools give you an on/off switch. Oversee gives you four controls instead: limit, preview, stop and undo.
An on/off switch next to Oversee's four controls, drawn in the Minia design system the app uses.
Each action gets a label that decides if it can run alone
Safety shouldn't depend on someone being quick enough to catch a mistake. Some things the agent just can't do on its own.
There are four labels. Steps you can undo just run. The practice run only previews. Important steps wait for your approval. Steps that can't be undone never run without you. Each label has an icon as well as a color.
The four labels, drawn in the Minia design system the app uses.
Practice run: seeing what will happen first
A practice run goes through the plan on a copy of the data and shows what would happen. Nothing real changes.
You see the same matches, flags and totals the real run would give, while they're still easy to change. The numbers are clearly marked as a forecast.
A practice run forecast, drawn in the Minia design system the app uses.
Stopping the agent partway
The stop button is on screen the whole time the agent works. Stopping is a normal action, not a panic button.
When you stop, Oversee shows what's done, what's in flight and what's still queued. Then you can carry on, give new instructions, or undo. Take away a permission and that step disappears from the plan, so the agent stops itself partway. Try it above: start the run, then press Stop.
What you see when you stop a run, drawn in the Minia design system the app uses.
Steps you can undo run alone, the rest wait for you
Being able to undo is what actually keeps this safe. A step that can't be undone stops and asks first, and shows what's at stake, not just a "confirm" button.
Most of an agent's work can be undone, so it can just run. The rare step that can't, like releasing a payment or emailing a customer, never runs without you. Oversee stops there, says what it's about to do and how much, and says plainly that it can't be undone.
The approval step, drawn in the Minia design system the app uses.
How the design changed from v1 to v7
Each version added one thing.
It started as a plan the agent just ran. Each version closed the gap between what the agent could do and what you could see, stop and undo. It ended with one rule: a step that can't be undone never runs without you.
-
1
Plan, then run
The agent made a plan and ran it. You could read the plan, but not see what it would do.
-
2
+ Practice run
Showed the matches, flags and totals on a copy of the data before anything changed.
-
3
+ Permissions
Each thing the agent can do became a switch. Moving money always needs approval.
-
4
+ Step by step
Replaced the spinner with a list of labeled steps you can follow as they run.
-
5
+ Stop button
Put stop on screen whenever the agent works, with a summary of what is done, in flight and queued.
-
6
+ Undo, where possible
Added undo for steps that can be taken back. It never claims it can unsend money.
-
7
+ Approval step · current
The one step that can't be undone stops and shows exactly what's at stake before you say yes.
The other three prototypes in the lab
Steer agrees the brief before the AI writes. Ground shows where each claim comes from. Recall shows what the AI remembers about you, and lets you delete it.