Solana Foundation
Decentralize, grow, and secure the Solana network.
Start new thread

Should AI agents have access to confidential transfers?

by•

Last May, launched , an open-source, real-time payment infrastructure for AI agents so they can discover and pay per call for any API.

There's now a proposal on the standard behind it that would encrypt the amount on-chain. In plain English, only the sender and recipient can see what actually moved, and nobody outside the transaction is required to check the numbers.

Would you let an agent spend on your behalf if you couldn't see how much?

125 views

Add a comment

Replies

Best
Privacy is important, but AI agents should never have unlimited spending power. Every confidential payment should still respect user-defined limits, permissions, and verifiable settlement.

 human-in-the-loop ftw

 Exactly. AI should accelerate decisions, not replace user authority. Human-in-the-loop plus verifiable settlement creates a much safer model for financial transactions.

Nah. Hide it from the public, not from me. If we built the agent and it’s spending our money, I need to see every payment.

"If we built the agent and it’s spending our money, I need to see every payment."

 and what's your opinion on commits? see this . are you more human-in-the-loop or agent-first?

 Great question. I’m agent-first. Let the agent do its thing. But once it’s spending real money or making changes you can’t undo (like deploying to a real cloud), I want budgets, guardrails, and a way to see exactly what it did.

 makes sense! (full disclosure: 100% agree with you) ✌️

This has actually been an ancient topic from the web3 community actually, the 402 error was designed and kept partially for that I recalled.

 til!

the visibility question isn't really about watching it happen in real time for me, it's whether i can reconstruct what happened after the fact if something looks wrong. if the amount's encrypted for everyone except sender and recipient, that also blocks whatever monitoring or reconciliation tooling i'd normally point at the transaction history, unless it's explicitly handed a view key. so it's not one settings toggle, it's deciding in advance every system that might need to audit this later and making sure it's on the allowlist before something breaks, not after

i dont think given the probabilistic nature of AI agents they should have unlimited spending power