Using the approval queue

How proposed actions reach you to approve or deny, and how the soft-commit window lets you intercept an automatic action before it lands.

Your home screen groups the agent's work by state: “Needs you” holds proposed actions waiting on your decision, “Happening now” shows supported automatic actions still inside their safety window, and separate groups cover completed or upcoming work.

An item reaches “Needs you” because it exceeds a configured threshold, belongs to a human-required category, or is part of a capability you have not enabled for autonomous handling. Completed eligible work retains a plain-language reason and status so the queue is a review surface, not just a list of buttons.

Each item in “Needs you” includes the proposal, supporting context, and Approve and Deny controls. A denial can require a reason, which is retained with the decision and can be included in an affected party's communication where the workflow calls for one.

Higher-stakes money actions can require an extra confirmation rather than a single tap. Whatever you decide is written to the activity record with the actor, timestamp, proposal, and outcome.

Actions that can't be taken back — moving money, posting to the ledger, sending an outbound notice — don't fire the instant the agent decides. They're held for a short soft-commit window (about 15 minutes) during which you can stop them. In "Happening now" each one shows a live countdown and a Stop control, and the same cancel link rides along in the push and SMS alerts, so you can intercept from your phone without even opening the app.

Lower-risk, easily reversible actions don't need a hold and simply run, since undoing them afterward is trivial.

There are two ways to walk something back. Cancel stops a pending action before it ever executes — available while it's still in its window. Undo reverses an action that has already committed, for the reversible ones. Both live on the agent activity feed, alongside a plain-language "why" for every decision the agent made.

Approval and denial records give you evidence for reviewing a threshold. Repeated approvals do not automatically widen an agent's authority, and repeated denials do not silently rewrite policy. Instead, review the pattern by action type, amount, property, and stated reason before changing a limit.

Keeping the policy change owner-controlled separates learning from permission: the record can show where a setting creates unnecessary review or too much risk, while the owner remains the person who changes the boundary.