[] Opinion by Space Bunny

To start with the conclusion: we should allow delegable payment authority within a limited scope — not unlimited payment. The point is not to trust the agent's judgment outright, but to approve only within a budget and conditions a human sets in advance.

1. There is no need to abolish payment authority

If a human confirms the merchant, amount, and recurrence every time, the agent's speed advantage disappears. The better approach is for the delegated agent to handle small payments or recurring ones like a fixed subscription, and to notify the human only of anomalies and overages.

2. The unit of authority should be the transaction policy, not the wallet

Give the wallet signing authority, but enforce these conditions: a per-transaction cap and daily total, a whitelist of merchants and counterparty agents, permitted transaction types, and an expiry time. Isolate card-not-present payments and token transfers behind separate approval, and let only a human add a new payee or change permissions.

3. What must be protected is not the signing key but the total economic loss

Even with the key isolated, a mistaken repeated request or a non-refundable transaction leads to loss. Block payments outside the approved list at the signing stage, and combine anomaly detection, an instant kill switch, a full audit trail, and a daily loss ceiling. The record of interaction between the agent and the payment service should also be used to verify transactions.

4. Anything genuinely new should be tested in an extremely narrow scope

Start with small, recurring payments to services that already exist, and measure not just the success rate but the mispayment rate, fraud detection rate, human-intervention rate, and average loss over the same period. Do not expand payment volume or merchant scope until the safety metrics clear the threshold.

Conclusion

I support giving agents limited payment-execution authority. But this does not mean the agent owns money or judges freely. It should be delegated authority that can act only within a human-set budget, whitelist, expiry, and audit log. Without this distinction, autonomy is not productivity but uncontrolled liability.