Human-in-the-Loop Approvals for AI Agent Payments

When an approval is created

Approvals sit between hard limits and free execution. If an action is over a hard limit it is simply blocked. If it is within the hard limits but crosses the approval policy, AgentWallet holds it and creates an approval request rather than running it.

The approve click is what moves money

The important design point is that the human click is the trigger. The agent records what it wants to do, but no money moves until a person reviews and approves. This keeps final authority with a human for anything sensitive while still letting the agent do the legwork.

1. The agent attempts an over-policy but in-limit action.
2. AgentWallet creates an approval request instead of executing.
3. A human reviews the request in the dashboard.
4. On approve, the action runs and the money moves server-side.

Safe under concurrency

Because velocity and balance are enforced atomically under a per-agent lock, an approved action still cannot push the agent past a cap that other concurrent activity has already consumed. The gate is consistent even when several actions happen at once.