The agent negotiated the deal in 340 milliseconds. Who enforced it?
An AI agent has just negotiated a purchase across four parties — seller, marketplace, delivery provider, affiliate — in under a second. No human was involved. No contract was signed. The question is not whether the AI can negotiate. It already can. The question is: who enforces the agreement?
Human trust infrastructure — contracts, courts, banks, customer support — works because it operates at human scale and human speed. An AI agent processing thousands of transactions per day with unknown counterparties globally, 24/7, cannot call Stripe's dispute line. The economics do not work. The latency does not work.
The AI does not need to move onto Radix. It operates on whatever infrastructure is fastest and cheapest. Radix is invoked only at the moment where shared trust between multiple parties is required — when no single party's database is sufficient.
The things that stay off-chain — reasoning, search, negotiation, logistics — are things where a single party's database is sufficient. The things that go on-chain — payment, permission verification, revenue split, budget enforcement — are things where multiple parties need a shared record that none of them controls.
The single on-chain settlement event enforces the full multi-party split atomically. This is not custom smart contract logic — it is the standard Radix manifest worktop pattern, human-readable and inspectable by any party.
Unlike an EVM transfer() call — opaque without a decompiler — a Radix manifest states in plain terms what assets moved, from where, to where, and under what conditions. Any party can read it. Any party can verify it.
An AI agent has no emotional attachment to a bank or payment provider. It evaluates infrastructure mathematically: cost, reliability, enforceability, availability. A neutral settlement layer with deterministic rules and no downtime scores better on all four metrics than any intermediary requiring bilateral trust.
The pre-authorised spending envelope — the buyer signs a subintent before the agent acts, not at execution time — is both faster and more trustless than real-time approval. The buyer's intent is cryptographically committed, not verbally agreed. The agent operates within it without returning for confirmation on each transaction.