Swipe to Approve: Mobile-First Agent Safety
Swipe to Approve: Mobile-First Agent Safety
You're at lunch. Your phone buzzes. Your AI agent wants to delete a database migration file and it's waiting for your approval. You glance at the notification, see the context, swipe right to approve, and go back to your sandwich. Total interruption: four seconds.
That's the experience we designed with SafeClaw's swipe-to-approve interface, and it's changed how our users interact with their agents.
The Problem with Desktop-Only Approvals
SafeClaw's escalation system works beautifully when you're sitting at your computer. An action gets flagged, a notification appears in the dashboard, you review it, click approve or deny. Simple.
But developers aren't always at their desks. They step away for coffee. They're in meetings. They're commuting. And when an agent is waiting for approval, every minute of delay is a minute of lost productivity. If approvals are too slow, teams either stop using escalation (dangerous) or switch to overly permissive policies (also dangerous).
We needed an approval interface that was as fast and available as a text message.
How It Works
When SafeClaw escalates an action, it sends a push notification to the SafeClaw mobile PWA. The notification includes a summary of the action — what the agent wants to do, which file or resource is involved, and why it was escalated (which policy rule triggered it).
Tapping the notification opens a compact approval card that shows:
- The action type and target (e.g., "Delete file: src/migrations/003.sql")
- The risk signals detected, if any
- The agent's recent action history for context
- The policy rule that triggered the escalation
From this card, you can:
- Swipe right to approve the action
- Swipe left to deny it
- Tap to expand the full details and see the complete action parameters
The swipe gesture was deliberate. We wanted approval to be a physical action — something that requires intention, not just a tap that could happen accidentally. The swipe direction is also configurable for users who prefer a different mapping.
Designing for Trust on a Small Screen
The biggest design challenge was showing enough context for a confident decision on a phone screen. A desktop dashboard has room for file diffs, full command strings, and session history. A phone notification does not.
We solved this with progressive disclosure. The notification itself shows the one-line summary. The approval card shows the essential context. And a single tap expands to the full details. Most approvals happen at the card level — users develop pattern recognition for common escalations and can approve or deny in seconds.
For truly complex escalations, we include a "Review on Desktop" button that keeps the action pending and sends a reminder to the dashboard. Not every decision needs to happen on mobile.
Security Considerations
Mobile approval introduces authentication concerns. We require biometric authentication (Face ID, fingerprint) or PIN before any swipe-to-approve action is processed. This prevents someone who picks up your unlocked phone from approving agent actions.
The communication between the SafeClaw daemon and the mobile PWA uses encrypted channels. Approval tokens are single-use and time-limited. A token that isn't used within 10 minutes expires and the escalation returns to pending.
The Impact
Since launching swipe-to-approve, our users report that median escalation response time dropped from 8 minutes to 23 seconds. That's the difference between an agent waiting awkwardly and an agent that barely pauses.
More importantly, teams are more willing to use aggressive escalation policies because the cost of escalation — developer interruption — is now much lower. More escalations, faster responses, safer agents.
Get Started
Swipe-to-approve works through SafeClaw's PWA, which you can install from the dashboard. Setup takes under a minute. Full instructions are in our documentation, and the implementation is on GitHub.
The best safety system is the one people actually use. By making approvals fast and mobile, we removed the biggest friction point in human-in-the-loop agent safety.