An open porcelain wallet beside a porcelain signet stamp with a violet glass handle

Agent wallets vs agent spend authority

A wallet holds money and enforces a balance. Authority decides what the money is for. When you need one, the other, or both.

Neltava Team · Updated

Two different jobs

When a team first lets an agent spend, the question sounds like one question: how do we let it pay without losing control? It is really two. One is about money: where it comes from, how the agent reaches it, and the hard ceiling on how much it can move. The other is about permission: whether a particular spend is something the agent was delegated to do.

Agent wallets — balances, virtual cards and accounts built for software to hold — answer the first. Spend authority answers the second. They are different layers, and most setups that go wrong treat one as if it were the other.

MCP / tool protocol

Can the agent call it?

Mandates / agent identity

Was the agent authorized to act?

Payment rail

Can the payment settle?

Spend limit

Is it within the number?

Neltava

Should this agent do this — under this delegated authority, for this purpose, right now?

A wallet is a payment rail. Authority is the layer that asks whether the payment should happen.

What a wallet decides

A wallet is excellent at the things a balance can know:

  • whether there is enough money;
  • whether a payment is under a card or wallet limit;
  • whether a merchant or address is on the card’s allowed list;
  • how much has moved this period.

These are hard limits on the money itself, and they are valuable precisely because they don’t depend on the agent behaving. A $500 prepaid balance can never lose $5,000.

What authority decides

A balance cannot know what the money was for. Authority adds the questions that need the agent’s mandate:

  • Purpose. Is this spend for the reason the budget was delegated? A B2B lead-generation budget buying a consumer campaign is within every limit and still wrong — see purpose-aware spending.
  • Sequences. Five small payments to one payee that add up to what a person should have approved.
  • People. Which spends need a person, who that person is, and the guarantee that the requester can’t approve itself — see human review.
  • Evidence. For every spend: what the agent asked, why, under which version of its authority, what was decided and what happened.
  • One policy across rails. Agents rarely spend from one place. Authority applies the same rules whether the payment goes through a card, an invoice, an API top-up or a wallet.

Side by side

Agent walletSpend authority
Core questionCan this agent pay, and from what balance?Should this agent make this payment?
Holds moneyYes — a balance, card or accountNo — never moves or holds funds
LimitsBalance, card limit, per-transaction and period capsPer-transaction, daily/monthly, windowed totals per payee or category, velocity
Who may be paidOften an allow/block list per card or walletAllow/block lists, trusted categories, payee identity
What it's for—A written purpose, checked on every spend; doubt goes to a person
People in the loopSometimes approval for large paymentsReview for doubtful spends, separation of duties, labels
EvidenceTransaction historyEvery decision with its authority version, trace and outcome, hash-chained
RailsIts ownAny — cards, invoices, APIs, prepaid balances, wallets
Testing a policy safelyUsually liveShadow Mode: decided and recorded, nothing blocked

Using both

The strongest setup uses each for what it is good at:

  1. The wallet holds a bounded amount. Its balance and limits are the last line: whatever else fails, the agent can’t move more than is in it.
  2. The agent asks Neltava before every payment. The spend is checked against the agent’s authority and purpose, and doubtful ones go to a person.
  3. The wallet pays only with a capability. In Enforce, an allowed spend carries a single-use capability bound to its payee, amount and currency. Whatever controls the wallet — your payment service, a signing service — consumes it before it pays. That is the payment adapter pattern.
  4. What the wallet reports becomes the outcome. Failed payments give their budget back; executed ones count at the amount that really moved.

Which one you need

A wallet alone is enough when…

the agent pays one kind of vendor, for one kind of thing, from a balance small enough that its limit is the whole policy.

You need authority when…

the agent chooses what to buy, from many payees, with a budget that belongs to someone else — and “under the limit” is not the same as “what we wanted”.

You need both when…

real money moves at volume. The wallet caps the worst case; authority decides the ordinary case, keeps people on the doubtful ones, and leaves a record you can defend.

When you don’t need Neltava

Start in Shadow Mode

See what your agents would have done — before anything is blocked.

Connect one agent through the SDK, MCP or one HTTP call. Shadow Mode is free: every spend is decided and recorded, nothing is blocked.

Start free in Shadow Mode →no card · one agent in minutes

Keep reading