Human-in-the-loop UI: approvals, permissions and autonomy
Human-in-the-loop UI lets agents act on their own where it's safe and ask where it isn't. The core patterns are an approval card for single actions, a permission request for access, an autonomy setting that decides when to ask, steering for running work, and a review step for changes.
Ask about the action, not in general
"The agent wants to continue" is not something people can judge. A good approval shows exactly what will happen – send this email to these people, charge this amount, delete these files – with the inputs visible and two clear answers. Keep it next to the work it belongs to.
- Approval Card – An agent asks before it acts – risk badge, preview, Approve and Deny with shortcuts, a pausing countdown, and a smooth morph into a resolved row
- Decision Inbox – Open decisions with two quick answers each – answered ones confirm and fold away
Permissions: scoped and reversible
When an agent needs access to a tool or data, ask once with a clear scope – read or write, this folder or all, this time or always – and make it easy to take back later.
- Permission Request – An agent asks for one specific right – who, why and exactly what will happen – and you allow it once, for an hour or always; then it folds into one line with the time left and Revoke
- Integration Settings – Settings card for a connected service – on/off switch, API key, validated site address, time zone and a save bar
Autonomy: decide when the agent asks
Not every action needs a person. Let people choose per kind of action: always ask, ask only when risky, or run on its own. Destructive actions like deleting or spending money should stay locked to ask first.
- Autonomy Dial – How much an agent may do on its own – a three-stop dial, and its real actions gliding between “Runs on its own” and “Asks you first”, with a lock for what always asks
Steering work that is already running
People notice problems while an agent works. Let them queue a note the agent picks up at its next step, or interrupt it, instead of stopping the whole run and starting over.
- Agent Steering – Steer an agent while it works – queue a note for the next step and watch it glide into the plan when the agent picks it up, or interrupt right away
- Agent Plan – An agent's step-by-step plan to review before it runs – rename, drag to reorder, remove or add steps, then watch each one run with live timers and retry from a failed step
Reviewing what changed
Before changes go live, show them as a diff and let people accept or reject parts instead of all or nothing. When agents queue many actions, a review workspace sorted by risk lets people go through them one at a time.
- Inline Diff – AI edits to prose shown inline – struck and added words you accept or reject one by one, with keys, a popover and undo
- Review Center – A focused workspace for what your agents want to do – the queue sorted by risk, one item at a time with the shaky spans marked, Approve, Edit, Reject and Skip on single keys, a stamp as each one leaves and the next one rising in
Questions
When should an AI agent ask for approval?
Before actions that are hard to undo or affect other people: sending messages, spending money, deleting data, publishing, or changing permissions. Reading and drafting usually don't need approval.
How do I avoid approval fatigue?
Ask about concrete actions, group similar ones, and let people raise the agent's autonomy for actions they trust – so they only see the decisions that matter.
Updated 2026-10-29