Should AI agents be allowed to hold private keys or credit cards?

x402 is making it much easier for agents to pay for APIs.

But most implementations still assume a trusted agent (often holding private keys or payment credentials).

Curious what people think the right authorization model looks like:

• full access (simplest, but risky)
• human-in-the-loop
• scoped permissions (limits, merchants, expiry)

If agents are going to operate autonomously, feels like this layer becomes critical.