Research and agent development
The AI Sandbox is FNA's test service in which AI agents act for banks against Sampo on simulated payments. It runs on the Sampo Simulator. Each agent has its own bank's payments and uses Sampo as a bank would in production.
Sampo coordinates across the banks in a session and gives each agent advice on its own payments only. An agent never receives another bank's payments.
The AI Sandbox runs the clock
The AI Sandbox runs a simulated RTGS and the clock. It runs a simulated day in turns, so agents can take as long as they need in each turn. The AI Sandbox closes a turn when every agent has finished, or when the turn's time limit passes.
Your simulation runs the clock
An external payment simulation runs its own clock, agents and RTGS. It reports to Sampo what its RTGS did and asks Sampo for advice at each step.
The agent reads its seat's status: the simulated time, its payments, its net debit and its funding checkpoint.
The agent reads Sampo's advice for this turn, which lists the payments to submit now. Advice lapses at the end of the turn.
The agent checks each payment against the funding checkpoint in the advice, and receives its headroom after the payment (field headroom of check_release). A negative value is a shortfall.
The agent submits payments to the simulated RTGS, on the advice or on its own judgement.
The agent sends new payment intents to Sampo (submit_instructions), and amends or cancels earlier ones.
The agent ends its turn. When the day closes, each agent reads its reports.
Agents connect over a remote Model Context Protocol (MCP) server or an HTTP API with an OpenAPI 3 contract. Both offer the same operations.
The MCP server includes guidance written for an agent that has read nothing else: what Sampo does, how a turn works, what a seat can see, and the rules that cause rejections. Each tool's description states when to use it, what it returns and how to fix its common errors.
Each agent authenticates with an API key for its seat. Invited users create keys in the web interface. A key is shown once and can be revoked at any time.
Example: an agent's turn, as MCP tool calls
get_status()
get_advice()
check_release(sampo_payment_id=8)
submit_payments(ids=[8, 11], request_id="t42-submit")
submit_instructions(instructions=[...], request_id="t42-send")
end_turn()The agent reads its status and Sampo's advice, checks its headroom after the payment for Payment 8 (check_release), submits Payments 8 and 11 to the simulated RTGS, sends new payment intents to Sampo (submit_instructions) and ends its turn. The documentation for invited users lists every operation, with its inputs and errors.
Five synthetic days of payment intents, 500 payments between five banks each, from FNA's research generator. The data is calibrated to published Fedwire structure and contains no bank data. Each row gives the sender bank, receiver bank, amount, priority, earliest time, own time, and when the bank sends the payment intent to Sampo. The files carry no Sampo results: no advice, no submission times with Sampo and no peaks with Sampo. Terms of use: CC BY 4.0.
Sessions run on synthetic payments. Do not upload real payment data.
The AI Sandbox holds payment data in the clear. Multi-party computation is part of the production design, not the AI Sandbox.
When a session expires, the AI Sandbox deletes its log, outputs and exports. Logs contain no amounts, references or bank names.
Access is by invitation, for research and agent development teams. Request access
Ask for AI Sandbox access, an evidence pilot on your own payments, or more information about Sampo. Submitting the form opens your email app with a message to phillip@fna.fi. You can edit it before you send it. This site sends and stores nothing.