CORTEXFLOW · FREE FOREVER

News

JPMorgan Just Let Two AI Agents Settle a Currency Trade — With No Human in the Loop

JPMorgan Just Let Two AI Agents Settle a Currency Trade — With No Human in the Loop
FIG. 01 — JPMorgan Just Let Two AI Agents Settle a Currency Trade — With No Human in the LoopSCALE 16:9

On November 17, 2026, JPMorgan Chase and South Korea’s Hyundai Card completed something no bank had tried in public before: two AI agents, one belonging to the bank and one to the client, carried out a foreign-exchange transaction between themselves — with no human in the loop. The trade was simulated, no real money moved, and the banks are careful to say so. But the part that matters for anyone building agents that touch money is not the trade. It is the box around it.

JPMorgan’s Kinexys unit built a control layer that confines each agent to predefined transaction parameters and automatically rejects anything outside them. “You are going to draw a box around the agent,” Wee Kee Toh, global head of Kinexys Insights, told CNBC. “The agent will never be able to do anything out of the box.” That sentence is the whole design pattern — and it is one you can steal for your own n8n workflows today, long before any bank lets your agent near a real payment rail.

What actually happened

The proof of concept ran as a corporate treasury scenario — the work of managing a company’s cash and currency across borders. Here is the sequence, as reported by CNBC on November 17 and demonstrated at the Singapore FinTech Festival on November 18:

  1. The client’s agent asked for liquidity. A Hyundai Card AI agent noticed the euro balance in its Kinexys blockchain deposit account needed topping up, checked balances and compared foreign-exchange rates across banks, then requested a trade.
  2. The bank’s agent checked the other agent, not just the trade. Kinexys’ AI agent verified the Hyundai Card agent’s identity and authority before doing anything with the request.
  3. The trade executed inside pre-set limits. The exchange and transfer ran automatically, and the account balance updated after processing.
  4. An out-of-box request was blocked. In the same demonstration, a transaction that would have dropped a pound account below £4.5 million hit a pre-set floor rule — and Kinexys’ control system rejected it outright.

Kinexys is JPMorgan’s blockchain business unit, the platform behind the bank’s cross-border transfers and real-time payments for corporate clients. JPMorgan calls the exercise the first private-blockchain test between autonomous bank and client agents for corporate treasury management. That “first” is the bank’s own characterization, not an independent verification — but the shape of the test is what counts: two agents from two different institutions, negotiating and settling, with the enforcement done by a layer neither agent controls.

The numbers around the platform give the test its weight. Kinexys has processed more than $4 trillion in transactions since launch, averaging about $7 billion a day. The next phase, targeted for the first half of 2027, adds more agents that negotiate and agree terms with each other before paying — not just executing instructions. JPMorgan has set no timeline for commercializing any of this.

The box is the product

Kinexys control box: a client agent's FX request rides a rail as a trade slip in a robot cart's basket and slides into JPMorgan Kinexys' control box. The box checks identity, a $5M ceiling and a £4.5M balance floor — the two dials on its top — and only then does the stamped trade ticket drop down the chute into the output tray. A request outside the box is stamped into the reject bin.

Fig 1 — The Kinexys control box. The trade slip enters the box; the two dials on top are the $5M ceiling and the £4.5M balance floor. Inside the box, identity, authority and limits are checked — the approved ticket drops to the tray, the out-of-box request is stamped into the reject bin. Readout: $5M ceiling · £4.5M floor · $4T Kinexys volume · ~$7B/day · 0 humans · simulated.

Read the coverage and a pattern jumps out: every executive quoted stresses the same thing. The technology was the easy part. Toh said building the agent-to-agent trade was “relatively straightforward compared with establishing governance and internal controls.” The hard problems are accountability and the risk of letting AI agents execute financial transactions without direct human oversight.

That is exactly backwards from how most agent projects are built. The usual order is: get the agent moving money, then bolt on limits later. JPMorgan’s order is: define the box first, then let the agent move inside it. The control layer sits between the agent and the payment system, and it enforces three things:

  • Identity. The bank’s agent verifies the counterparty agent’s identity and authority before the trade is even considered. An agent cannot just show up with a request.
  • Ceiling. Each agent carries a maximum transaction size. An agent authorized to transact up to $5 million is automatically blocked from executing anything above that threshold — the block is mechanical, not a suggestion the agent can talk its way past.
  • Floor. Account-level rules protect balances. The £4.5 million pound-floor rejection in the demo is the clearest example: the rule fired on the numbers, not on anyone’s judgement in the moment.

None of this is new technology. It is old technology — limits, allowlists, hard rejects — placed where the agent cannot reach it. That placement is the entire idea. If the agent can edit its own limits, there is no box.

Why this lands the week it does

This is the third time in four days that the money rails have made the news, and the three stories form a single argument.

On November 16, Chase went down on a software bug in on-premise hardware, slowing transactions at five banks plus Zelle — a reminder that the rails themselves fail, and that an agent mid-payment has nowhere to go when they do. On November 24, Wells Fargo’s updated Online Access Agreement takes effect, deeming anything an agent does under your login as initiated and authorized by you, at your own risk. And now JPMorgan shows the other side of the same coin: the bank that lets agents transact is the one building the hardest walls around them.

The through-line for builders: the institutions are converging on the same rule — the agent acts, but a separate layer decides what it is allowed to do. Wells Fargo puts the liability on you. JPMorgan puts the enforcement in its own control layer. Either way, an agent with no box around it is now a liability story waiting to happen.

Steal the pattern: draw the box in n8n

You do not need a blockchain platform to copy this. The Kinexys control layer maps directly onto five nodes you can build in n8n this afternoon. The rule that makes it work is the same one JPMorgan used: the limits live outside the agent’s reach.

Step 1 - Put the limits in a place the agent cannot edit

Store the authority limits — maximum transaction size, allowed accounts, daily caps — in a source the agent workflow reads but never writes: a static table, a locked Google Sheet, or n8n’s own credentials store. If the AI Agent node can see the limits in its own prompt and nothing else enforces them, the box is made of paper.

Step 2 - Verify the counterparty before the trade

Before any money node runs, add a verification step: an HTTP Request or Code node that checks the requesting agent or user against an allowlist of identities and scopes. This is the Kinexys identity check, shrunk to a lookup table. Unknown identity, expired scope, or a scope that does not cover this transaction type: stop here, no trade.

Step 3 - Enforce the ceiling with an IF node, not a prompt

After verification, route the amount through an IF node: amount is less than or equal to the stored ceiling, and the post-trade balance stays above the stored floor. Both conditions must pass. A prompt instruction (“never spend more than $5,000”) is a suggestion; an IF node is a wall. The Kinexys demo blocked the £4.5 million floor breach mechanically — your workflow should do the same.

Step 4 - Approve anything the box does not cover

Anything outside the pre-set box — a new payee, a larger amount, a currency you have not whitelisted — routes to a human approval step (Telegram, Slack, or email with an approve/reject button) and the workflow waits. This is the same approval-gate pattern behind MoonPay’s PayBox, where the buyer approves every payment. The box covers the routine; a human covers everything else.

Step 5 - Log every decision where the agent cannot rewrite it

Write every attempt — approved, blocked, and rejected — to an append-only log outside the workflow: a separate sheet, a database table, or a Telegram channel. Include the amount, the counterparty, which limit fired, and the timestamp. When something goes wrong, this is your proof of what the agent was and was not allowed to do. Wells Fargo’s November 24 terms make keeping proof of revocation a practical necessity, not a nicety — see our coverage of the liability shift.

Warning: Never let the agent node itself be the thing that checks its own limits. If the model decides whether it is inside the box, there is no box — the check must run in a node the model cannot influence, on data the model cannot edit.

Honest limits

Keep the story in proportion. This was a proof of concept, not a product: simulated transfers between hypothetical overseas affiliates, no real money moved, and no commercialization date. The “first” claim comes from JPMorgan itself. The next phase — agents negotiating and settling with each other in the first half of 2027 — is a testing target, not a launch. And the governance questions Toh himself flagged — who is accountable when two agents agree a bad trade, how authority is enforced across institutions — are unsolved. Hyundai Card says it plans to expand the collaboration into new areas, including in-car payments, and the broader Hyundai Motor Group operations were one reason JPMorgan picked Hyundai Card as its partner, but no specific additional projects have been announced.

The lesson is not “banks have solved agent payments.” It is that the biggest bank in the world, when it finally let agents move money, spent more effort on the box than on the agent. Build your box first too.

FAQ

Did JPMorgan let AI agents move real money?

No. The November 17 proof of concept used simulated transfers between hypothetical overseas affiliates. No real funds moved, and JPMorgan has set no timeline for commercializing the technology.

What is Kinexys?

Kinexys is JPMorgan’s blockchain business unit, providing cross-border fund transfers and real-time payments for corporate clients. It has processed more than $4 trillion in transactions since launch, averaging about $7 billion a day.

What is the “box around the agent”?

A control layer that confines each agent to predefined transaction parameters — identity verification, a maximum transaction size, and account-level floors — and automatically rejects anything outside them. JPMorgan’s framing, via Wee Kee Toh: “The agent will never be able to do anything out of the box.”

When will agents negotiate trades with each other?

JPMorgan and Hyundai Card target the first half of 2027 for a more complex multi-agent testing phase, where agents negotiate and settle with each other. That is a testing target, not a commercial launch date.

How do I apply this to my n8n workflows?

Store limits where the agent cannot edit them, verify the counterparty against an allowlist, enforce ceilings and floors with IF nodes (not prompt instructions), route anything outside the box to human approval, and log every decision to an append-only store. The full five-step pattern is in the section above.

Key Takeaways

  1. JPMorgan and Hyundai Card ran the first bank-to-client agent FX trade — simulated, no real money, reported by CNBC on November 17, 2026 and demoed at the Singapore FinTech Festival on November 18.
  2. The control layer is the story, not the trade — identity checks, a $5 million transaction ceiling, and a £4.5 million balance floor, all enforced by a layer the agents cannot touch.
  3. Governance, not technology, is the gating problem — JPMorgan’s own words; the next phase (multi-agent negotiation, first half of 2027) is a testing target with no commercial date.
  4. Steal the pattern in n8n: draw the box first — limits in a read-only store, counterparty verification, IF-node enforcement, human approval for anything outside the box, append-only logging.
  5. This is the third money-rail story in four days — the Chase outage on November 16 and Wells Fargo’s November 24 liability terms point the same way: an agent with no box around it is a liability story waiting to happen.

Next: read how the Chase outage took Zelle down with it and what Wells Fargo’s November 24 terms mean for your agent’s liability — then audit one of your own money-moving workflows against the five-step box pattern above.

DOC. CF-RSS-001 · FORMAT: XML

Want more like this?

Practical tutorials on AI agents, automation, and chatbots — straight to your reader.

Subscribe via RSS

Doc. CF-NL-001 · Dispatch

Get the next build in your inbox

One practical email per new post. No spam, unsubscribe anytime.

One email per new post. No spam. Unsubscribe anytime.