Singapore's AI Agent Sandbox Has One Nonnegotiable Rule: Prove the Bot Had Permission
Singapore plans to test whether banks and payment providers can verify an AI agent's authority before it moves stablecoins. The pilot is less about agentic commerce than about a problem nobody has solved: when a machine spends your money, whose signature is on the transaction?
Who's liable when an AI agent moves money it shouldn't have? That's the question Singapore's planned payments sandbox is really built around, and it's the only part of the exercise that matters.
The Monetary Authority of Singapore wants a controlled environment where banks and payment providers can test how they verify an agent's authority before a payment in digital money goes through. In practice, that means stablecoins, and in practice, it means the boring, load-bearing plumbing that decides whether an autonomous program can spend your balance.
The Test, Stripped Down
Details are thin. There's no published start date, no confirmed participant list, and no scope document that spells out which stablecoins or which rails get tested first. Spokespeople didn't immediately respond to a request for comment. What's clear is the sequencing. The sandbox is being framed as an authorization test, not a settlement test, which tells you where the regulator thinks the risk sits.
Look at the stack Singapore has already assembled. Project Guardian has been running industry pilots since 2022 on tokenized assets and digital money. The stablecoin framework finalized in August 2023 requires issuers to hold at least 100 percent of reserves in cash or short-dated government securities, with base capital of S$1 million or half of annual operating expenses, whichever is higher. Amendments to the Payment Services Act that took effect in April 2024 extended rules to custody and transmission of digital payment tokens.
So this sandbox isn't a first step. It's the last mile. The stablecoin issuer rules are written. The custody rules are written. What's missing is the piece in the middle, which is a verifiable answer to the question of whether the thing clicking buy actually had the right to.
Three layers, and only two of them exist yet.
Why Agent Payments Break the Chain
Every authorization model in retail finance assumes a human at the edge. Card networks lean on chargebacks. Banks lean on one-time passwords and device fingerprints. Exchanges lean on withdrawal allowlists. Onboarding gathers know-your-customer data on a person or a company, and the account is bound to that identity.
An agent isn't the customer. It's a delegate. And there's no clean registry anywhere on earth that says what a specific agent is permitted to do on behalf of a specific principal at a specific moment.
Stablecoin rails make that gap expensive. A transfer settles in seconds and finality is final. The reversal window that card networks built their dispute machinery around doesn't exist. If a treasury agent misreads a decimal and wires 500,000 USDC instead of 50,000, there's no chargeback. There's a phone call and a hopeful email.
That's the fault lines here. The industry has spent two years building agentic commerce rails, from Visa's Intelligent Commerce to Mastercard's Agent Pay to Google's AP2 and Coinbase's x402 specification, and most of that work optimizes for the happy path. Singapore is asking about the unhappy one.
What the Banks Want to See
According to two people familiar with the negotiations, the institutions circling this sandbox want a signed delegation credential more than they want faster payments. That means a cryptographic artifact that names the principal, scopes the agent to per-transaction and per-day limits, restricts counterparties, and can be revoked in near real time. They also want an audit trail that maps back to a natural person or a legal entity, because compliance teams can't file a suspicious transaction report against a model version.
None of that's exotic technology. Scoped tokens and verifiable credentials have existed for years. The hard part is governance. Who issues the credential? Who revokes it when an agent is compromised? Does a bank accept a credential issued by a third party it doesn't supervise?
That last question is the whole ballgame. Reading the policy tealeaves, this looks less like a technology pilot and more like a liability negotiation dressed in sandbox clothing. If the bank eats the loss on an unauthorized agent payment, agents get tiny caps and the product never ships. If the corporate client eats it, adoption is slow but real. If an insurer sits in between, you get a market.
The calculus for MAS is straightforward. Singapore already regulates stablecoin issuers tightly, and it has a payments law broad enough to cover token transmission. Running the authorization test onshore lets the regulator shape the credential standard before a US or European consortium sets one by default.
What to Watch
The participant list is the first signal. If the sandbox pulls in two or three major banks plus a licensed issuer such as StraitsX, the output is likely to be a reference design. If it's mostly fintechs and a payments processor, expect bilateral APIs and no shared standard, which would be a waste of everyone's time.
Second, watch whether MAS publishes a consultation paper on agent delegation before the pilot concludes. Regulators that intend to write rules tend to consult early. Regulators that intend to study forever don't.
Third, watch the timing around Singapore Fintech Festival in November. That's the natural venue for a progress announcement, and MAS has used the event before to convert pilots into policy.
And fourth, watch the stablecoin issuers themselves. Reserve rules are settled, but redemption mechanics at machine speed aren't. An agent that can mint and redeem on demand is a different risk animal than one that can only spend.
The question now is whether a sandbox in one city-state can produce an authorization standard the rest of the world borrows, or whether the industry keeps shipping agent payments with a terms-of-service checkbox and a prayer. My read: the banks will show up, they'll demand the credential, and they'll walk if the liability sits with them.
That's not a technology problem. That's a contract problem, and it's about to get drafted.
Explore More
Key Terms Explained
An autonomous program that can perceive on-chain data, make decisions using machine learning models, and execute blockchain transactions without human intervention.
Coinbase's Layer 2 blockchain built on the OP Stack (Optimism's technology).
Following the laws and regulations that apply to financial activities, including crypto.
Who holds and controls your crypto assets.