Step-up is how Glide hands control back to you the human at exactly the right moment. The agent does the work, finds the next thing to do, computes that it crosses your envelope threshold, and pauses. You get a notification. You approve in the dashboard or app. The agent continues.Documentation Index
Fetch the complete documentation index at: https://glide-9da73dea.mintlify.app/llms.txt
Use this file to discover all available pages before exploring further.
When step-up triggers
Any of:- The transaction amount exceeds the envelope’s step-up threshold.
- The counterparty isn’t in your allowlist (and the skill template treats out-of-allowlist as step-up).
- The transaction velocity would exceed the rolling-window cap.
- The risk model returns a
flagverdict. - The agent is requesting a sensitive write (e.g., adding to the beneficiary allowlist, changing the policy envelope itself).
What you see
The dashboard shows a step-up card with:- Which agent is asking.
- Which tool it wants to call.
- The amount and counterparty (or a redacted preview if it’s a non-money tool).
- The natural-language explanation — an LLM-generated summary of what the agent’s about to do, in plain English.
- A diff view — before/after state if the action would change something.
- The before-and-after policy — if the agent is asking to change envelope settings.
How to approve
| Surface | How |
|---|---|
| Phone push | Face ID or fingerprint, directly from the lock screen |
| Mobile app | Tap the step-up card, biometric or passcode |
| Web dashboard | Tap Approve, then your passkey or two-factor method |
How step-up is bound to a single call
Every step-up generates a single-use sigil — a one-time token that authorizes the specific tool call you approved. Once consumed, the sigil is invalidated. Replay attacks can’t use the same sigil to approve a different call. If the agent retries the same tool call (e.g., the network blipped between approval and broadcast), it has to ask for a fresh sigil. This is intentional: every successful tool call is paired with exactly one human approval.URL-mode elicitation
Some runtimes can’t render step-up prompts inline (older chat surfaces, voice-only interfaces, headless agents). In those cases, the agent gets back a step-up URL — a link you open in the dashboard or mobile app, which surfaces the same approval card. You’ll see this most commonly in:- Voice-mode agents (Apple Intelligence summary mode, ChatGPT Voice).
- IDE-embedded coding agents that don’t have a chat-style approval UI.
- Server-side agent frameworks that don’t have a user attention surface.