Skip to navigation

Multiple wallets

Give a customer separate Agent Wallets for different spending purposes.
View as Markdown

A customer can have multiple Agent Wallets, each with its own balance, spending controls and funding addresses. Separate wallets let customers organise money by purpose and give their AI agents spending access to a specific balance without treating all of the customer’s funds as one unrestricted pool.

For example, a customer might keep separate wallets for travel, office expenses and research. Funding one wallet does not increase another wallet’s balance, and a payment made from one wallet does not draw funds from another. The separation remains in place even though all of the wallets belong to the same customer.

Use names customers understand

Wallet names are part of how customers communicate spending intent to their AI agents. A customer should be able to say something such as “Book the flight from Travel” and have Travel refer unambiguously to the intended wallet.

Choose names that describe the purpose of the money rather than internal identifiers that only make sense inside your application.

ClearAvoid
TravelWallet 1
OfficeWallet 2
ResearchOther
GroceriesMain 2

Names should also be unique within the customer because two wallets with the same customer-facing name would make it impossible for an AI agent to determine which balance the customer intended to use.

Wallet selection

The amount of information required from the agent depends on how many wallets the customer has and whether the operation can move money.

SituationBehaviour
Customer has one walletThe wallet can be used without requiring the customer to distinguish between several options.
Customer has several walletsA wallet must be identified when the operation will move money.
Wallet is not identifiedThe payment is not completed by silently selecting another wallet.
Wallet name is unknownThe payment is refused rather than falling back to a default wallet.
Balance is requested without a walletBalances can be returned across the customer’s wallets because no money is being moved.

The important distinction is between reading information and moving money. Returning several balances gives the agent more information but does not change the customer’s funds. Choosing a wallet for a payment changes which balance is spent, so Mitu leaves that decision with the customer rather than making an assumption on their behalf.

Preserve the customer’s choice

Suppose a customer has Travel and Office wallets and asks an AI agent to pay for a hotel from Travel. If the Travel wallet does not have enough money, the agent should not silently use the Office wallet simply because it has sufficient funds.

Those balances were separated intentionally. Moving to another wallet would change the customer’s instruction and could spend money reserved for a completely different purpose. Instead, the payment should stop and leave the customer to decide whether to fund Travel, choose Office explicitly or take another action.

The same rule applies to an unrecognised wallet name. A typo or a wallet name invented by the model should resolve to no wallet rather than whichever existing name appears closest.

Wallet selection is a safety boundary

Requiring a valid wallet choice also reduces the ability of untrusted content to redirect spending. An AI agent should not be able to move a payment to another one of the customer’s wallets simply because content it processed suggested doing so.

Design wallets around purposes

Create wallets for meaningful spending purposes rather than individual transactions. A Travel wallet can remain useful across flights, hotels and other travel expenses, while a wallet created for a single booking is harder for both the customer and their AI agent to understand or reuse.

Customers who only need one spending balance should generally have one wallet. Multiple wallets become useful when there is a genuine reason to separate funds, controls or purposes. As the number of wallets grows, names become more important because both the customer and their AI agent need to distinguish between them reliably.