Kakao Pay, the payments arm of Korea’s Kakao group, said on September 15 that it has completed a proof of concept for “agentic payment” — a system in which AI agents autonomously transact on a user’s behalf, settling in stablecoins. What makes the test notable is its directionality: the agents both bought goods or data and received settlement as sellers, a two-way flow that most agent-payment pilots, which focus on outbound spending, have not attempted. The announcement lands amid a burst of agent-commerce activity in Korea, including SK Telecom’s proposal for a human-approval layer over agent transactions.
What the PoC Actually Tests
According to Kakao Pay, the proof of concept covered a full transaction loop: an AI agent identifies something to buy — a product or a piece of data — negotiates or accepts the price, executes the payment in a stablecoin, and, on the other side, a seller-side agent receives and settles the proceeds. That second half matters. Most agentic-payment work to date, including the x402 protocol’s HTTP-native payment flow and Coinbase’s Agent Payments kit, has concentrated on the spend side: an agent pays a machine-readable invoice for an API call or a service. Two-way settlement — agents as economic counterparties in both directions — introduces harder problems around reconciliation, refunds, and dispute handling that card networks solve with chargebacks and that stablecoin rails currently solve poorly or not at all.
Why Stablecoins Are the Settlement Layer Here
Kakao Pay’s choice of stablecoins for the PoC is consistent with the technical constraints of machine-initiated payments. Agents need programmable, instantly settleable value with no card-association intermediation, and stablecoins on public chains provide that natively. It also fits a pattern we have covered repeatedly: stablecoin-native rails like x402 and Coinbase Agent Payments were designed around machine-to-machine pricing and payment, while card rails were not. The open question for a company like Kakao Pay — a licensed financial institution operating under Korean regulation — is which stablecoins, which chain, and under what custody arrangement. The announcement did not specify these details, which is a significant gap for anyone assessing production readiness. Korea will also begin taxing crypto transactions, including stablecoin spending, next year, with agent-initiated payments explicitly in scope — a compliance burden any production system must price in.
The Approval Layer: SK Telecom’s RCS Proposal
A separate but related development frames the Korean approach. SK Telecom proposed at a GSMA RCS Group standardization meeting a framework requiring users to review order details and payment amounts before an AI agent transaction is finalized — using RCS messaging as the confirmation channel. The implication is that Korean infrastructure players are converging on a model where agents negotiate and execute, but a human sign-off step sits at the point of payment. That is a different architecture from fully autonomous agent spending, and it directly addresses the trust problem that has slowed agent commerce: users will not hand spending authority to opaque software without a reviewable approval moment. Whether that approval step is inserted into every transaction or only above a threshold is the kind of design detail that will determine whether the model scales.
What This Means for the Agent-Payments Landscape
Kakao Pay is the latest non-crypto payments incumbent to move from watching agent commerce to testing it, following Ant International’s protocol launch and the card networks’ verification standards work. The Korean cluster is distinctive in combining three elements: stablecoin settlement, two-way agent transaction flows, and a telecom-proposed human-approval channel. If the PoC graduates to production, Kakao Pay’s user base — tens of millions of Korean consumers already using it for daily payments — would represent one of the largest real-world tests of stablecoin-settled agent commerce anywhere. The unknowns remain custody, chain selection, the tax treatment kicking in next year, and whether a two-way settlement model can handle disputes without a chargeback equivalent.