Where should authorization live in an AI agent payment stack?

AI agents are starting to use wallets and real payment infrastructure. Being able to make a payment is not the same as being authorized to make that payment. With x402Shield, authorization happens before signing and can enforce limits for amount, recipient, asset, network, wallet scope, budgets, approvals, replay protection and request binding. I am curious how teams building agent wallets, x402 infrastructure and autonomous commerce systems approach this problem. Should authorization live inside the wallet, inside the application, or in a separate layer before signing?