Aditi Mehta

AI agents can already browse and check out, but nobody trusts them with their money: one hidden fee or one line of small print aimed at the agent can make it overspend. PactPay × Meta Muse lets an agent shop while a fixed gate keeps the final say. You speak or type what you want. Muse Voice Transcribe hears you and Muse Spark drafts editable rules; you approve them once and they're hashed and locked. A Muse agent then shops by calling tools live: it reads your rules, searches real stores, ranks offers and asks for a PayPal hold, and PactPay refuses anything outside your limits. Everything runs as one workflow on one page. Before money moves, the checkout page is read twice (two different Meta models when both are configured, one of them open weights). Fixed code checks that the readings agree, that the itemized lines add up to the total, and that no text is aimed at an AI. The verdict is stored on the order and the settlement gate reads it, so a "Do not pay" blocks the capture and voids the hold even if the person approved it. The model reads. It never decides. Any other agent can plug in too: a limited, expiring agent key, an MCP server and an HTTP gateway, with every abuse refused (`pnpm agent:redteam`).