Approvals
When an agent wants to do something it is not allowed to do on its own, it stops and asks. Those questions collect in one place.
Choose APPROVALS along the bottom of the window. When something is waiting, the button carries an orange badge with the number of outstanding requests, so you do not have to go looking.

The header reads APPROVALS — 0 PENDING, counting up as requests arrive. With nothing waiting it says plainly:
Nothing awaiting your decision.
What a request looks like
Each request is a card. Across the top it carries a label saying what kind of
request it is, a one-line summary of what is being asked, and how long it has
been waiting — 12m ago, 3h ago, 2d ago.
If the request came out of a conversation, the card says which one.
Underneath, most kinds show a full card with the detail and the buttons to decide. A few show the summary line only.
The kinds of request
| Label | What is being asked |
|---|---|
| curriculum-ratification | An agent has been taught something and wants you to sign it off |
| curriculum-failure-choice | Teaching went wrong; you choose what happens next |
| skill-acquisition | An agent found a capability it would like installed |
| spot-check | You are being asked to grade a piece of work blind |
| graduation-trust-proposal | An agent has earned a case for more freedom |
All five come out of teaching agents, which is covered in Teaching an agent.
On a skill-acquisition card there is also a DISMISS button, labelled with exactly what it does:
DISMISS (installs nothing, rejects nothing — just closes this request)
Use it when you do not want to decide yet. Rejecting and dismissing are different things, and the button says so to stop you doing one when you meant the other.
The label that matters most
Some cards carry a red warning:
SESSION-SCOPED — dies with this run, not resumable
This means the request belongs to a task that is currently running. If you leave it, it does not wait for you. When that run ends, the request is gone and whatever it was asking about does not happen. You cannot come back to it later.
Treat a session-scoped request as something to answer now. Everything else can sit in the queue safely.
Decisions you have already made
SHOW RESOLVED switches the list to include requests you have already
answered. Resolved cards are dimmed and labelled · RESOLVED.
This is worth looking at every so often. It is the honest record of what you have agreed to, and it is easy to approve things one at a time without noticing the pattern they add up to.
REFRESH reloads the list.
What never reaches this queue
Some actions do not become an approval request because the app refuses them outright, no matter who asks or how much the agent has been trusted.
An agent can never run a command as an administrator. It is refused in the code before any approval prompt could be shown, so there is nothing for you to say yes to. When a job genuinely needs one, the agent is required to write out the command, explain what it does, and hand it to you to run yourself.
Three more categories always stop for a human and can never be set to run automatically:
- deleting, overwriting or bulk-changing things on real outside services
- payments, banking and anything that moves money
- creating accounts on public platforms in your name
Working with agents safely explains the thinking behind those rules, and Governance and the constitution shows where to read them inside the app.
A habit worth forming
Read the summary before you decide. A card is short on purpose, and the temptation is to approve it the way you accept a cookie banner.
The whole point of supervised mode is the few seconds you spend here. If you find yourself approving without reading, that is a sign you should either promote the agent deliberately — see Your first agent — or narrow what you are asking it to do.
