Scheduling the money: job shops, Hungarian matching and AI agents in payments
Every Friday afternoon, a treasury team solves a scheduling problem and a matching problem by hand. AI agents can frame both problems. Algorithms from the 1950s and 1960s already solve them.
By Pavel Nikolov, November 8, 2026On an average day in 2025, banks settled $2.014 trillion through CHIPS, The Clearing House's large-value dollar system, and funded it with about $96 billion. Under plain real-time gross settlement, the same payments would have needed an estimated $442 billion (The Clearing House, April 2026). The gap is an algorithm. Payments wait in a queue, and the CHIPS engine keeps choosing which ones to release, alone or offset against others, so each dollar of funding settled about $26 of value.
That is a scheduling decision, made continuously. Most other money movement still runs on the same kind of decision, made by people with spreadsheets and a calendar of cut-off times.
A treasury team on a Friday afternoon is solving a scheduling problem. Which payments go first, through which rail, funded from which account, before which window closes? A reconciliation team on Monday morning is solving a matching problem: which bank statement line belongs to which invoice, and which ones belong to nothing.
AI agents are now being handed parts of this work. Adyen launched Adyen Agentic on June 16, 2026, Google's AP2 lets an agent present a signed mandate at checkout, and the Machine Payments Protocol makes payment the authentication step for API calls (Fintech Wrap Up covered all three in its deep dives). Agents that buy things will soon be agents that move money between accounts, rails and currencies.
Language models frame
An agent can read “get the suppliers paid before the weekend without dipping below our buffer” and turn it into variables, costs and constraints. It can read a mandate, price a route and notice that a rail closes in 30 minutes.
Solvers decide
Job shop scheduling has had a standard benchmark since 1963. The Hungarian algorithm dates from 1955. Both return a plan and a measure of how good it is, which a language model guessing at the answer cannot.
I think the useful design for payment agents is that old pairing. Part 2 treats money movement as a job shop, with rails as machines and cut-offs as deadlines. Part 3 uses the Hungarian algorithm to decide who pays what, with a worked Friday-afternoon example. Part 4 shows how an agent frames the problem, hands it to a solver and runs the plan. Part 5 covers where this breaks, including the Taleb problem with optimizing away every buffer.
The rails, as a scheduler sees them
Every rail is a machine with opening hours, a capacity and a processing time. Their calendars do not line up, and the gaps between them are where most of the scheduling work sits.
| Rail | Open | Cut-off or limit | How it settles |
|---|---|---|---|
| Fedwire Funds | 9 p.m. ET the day before to 7 p.m. ET, Monday to Friday | Customer transfers close 6:45 p.m. ET | Gross, final, central bank money |
| CHIPS | About 21 hours per business day | Prefunded balance; payments queue until the algorithm releases them | Continuous netting, final on release |
| FedNow | 24/7/365 | $10 million network limit since November 12, 2025 | Gross, instant |
| T2, the euro RTGS | 22.5 hours, 19:30 to 18:00 CET, Monday to Friday | Customer payments 17:00 CET | Gross, final, central bank money |
| SEPA Instant | 24/7/365 | Funds in under 10 seconds; euro-area PSPs must send since October 2025 | Instant |
| Tokenized deposits (JPM Coin on Kinexys) | 24/7 | Limited to the bank's client network | Atomic, on a permissioned ledger |
| Public stablecoins (USDC, USDT) | 24/7 | On- and off-ramp capacity, conversion spread | Depends on the chain's finality |
The weekend shows the problem most clearly. From Friday 7 p.m. ET to Sunday 9 p.m. ET, Fedwire is closed, while FedNow, SEPA Instant and every stablecoin keep running. The Federal Reserve will open Fedwire on Sundays and weekday holidays in 2028 or 2029, and the ECB's 2026 consultation sets a path toward a near 24/5 T2.
Tokenized money removes the window and adds other limits. Corporate clients already use JPM Coin for overnight liquidity moves, and can pay between their own accounts in two countries on a Saturday (Fintech Wrap Up's deep dive on JPMorgan's payments strategy). A stablecoin payout still has to land in a bank account at the other end, through an off-ramp partner with its own limits and hours.
So the inputs to every Friday decision are calendars, capacities, costs and deadlines. That is the vocabulary of a job shop.
Continue to Part 2: Money movement as a job shop