JPMorgan Tests AI Agents in a Simulated Foreign-Exchange Deal

JPMorgan and Hyundai Card have tested AI agents in a simulated foreign-exchange deal, letting software carry out FX transactions without direct human intervention. The proof-of-concept, reported on November 18, is one of the clearest signs yet that banks are moving from talking about AI agents to letting them do real work.
The test ran on JPMorgan’s Kinexys blockchain platform. AI agents handled the steps of a foreign-exchange transaction in a simulated setting — a task that normally passes through several people and systems before it is done.
What the test involved
In the proof-of-concept, the agents carried out simulated foreign-exchange transactions on behalf of the two firms, according to CNBC. Hyundai Card, the South Korean card company, took part alongside JPMorgan.
Nothing here was live trading. The point was to see whether agents could follow the rules of a complex financial workflow — checking terms, executing steps in order and handling the handoffs — without a person stepping in at each stage.
Why banks are testing agents now
Foreign-exchange deals are a good test case. They involve several steps, strict rules and more than one party, so they show up the weak points of an AI system fast. If agents can handle FX, they can plausibly handle other multi-step jobs in banking too.
JPMorgan has been building out its blockchain and payments work for years, and Kinexys is the platform where much of it now sits. Putting agents on top of that platform is the next step: not just moving money on a ledger, but letting software run the deal itself. A write-up of the test is also available from Traders Union.
What comes next
A simulated deal is not a live one. Regulators, risk teams and auditors will all want to see how agents behave when something goes wrong — a bad price, a missing approval, a system outage — before real money is involved.
Still, the direction is clear. Banks are no longer asking whether AI agents can do financial work. They are asking how much of it they can be trusted with, and what guardrails need to be in place first. For a wider look at how AI agents are being governed elsewhere, see United Airlines’ decision to block AI agents and our report on the Linux Foundation Agent Name Service.
