← Go back

Guide // MandateGuard

How do daily spending caps stop AI agents leaking money?

By the Ingenious AI team · Published 2026-11-14 · Reviewed 2026-11-14

▸ Run the free live demo See pricing

Free covers 1 agent and 1 policy with the full authorize flow, signed receipts and the approval inbox — no card required to start.

Short answer

A daily cap bounds the total an agent may spend across one day, no matter how many individually legal purchases it makes. It is the control for the slow shapes of loss: retries after tool errors, loops that re-buy the same item, and gradual scope creep made of small charges that each pass a per-transaction cap. In MandateGuard the per-day cap is checked on every authorize call against the day's allowed spend so far. Requests that would push the day over the cap are denied with the reason recorded, and the attempts stay visible in the receipt chain.

What is a per-day cap?

MandateGuard's per-day cap is the maximum total USD an agent's policy allows in a day. Allowed decisions accumulate against it in real time. A request that fits the per-task cap but would exceed the remaining daily room is denied. Denied attempts do not consume budget — they create receipts. The cap resets with the new day, and the chain keeps the full history so daily patterns can be audited long after.

The losses that only a daily cap can see

A per-transaction cap inspects trees; a daily cap inspects the forest. Consider an agent that buys an 18-dollar data batch, gets a malformed response, and retries — eight times across an afternoon. Every request is small, allowlisted and plausible. Without a daily bound, the only limit is the agent's persistence. The same shape appears when two workflows unknowingly buy overlapping data, or when an agent compensates for a failing tool by purchasing from a second source. None of these is one bad decision; all of them are one bad day. The daily cap is the control shaped for aggregate behaviour, and its denials are often the first reliable signal that the agent, not the market, has a problem.

How MandateGuard counts the day

Counting is decision-based, not statement-based. Each ALLOW adds its amount to the day's total at the moment the engine answers, so the check on the next request already reflects it — there is no window where parallel requests can each see yesterday's total and jointly overshoot. Amounts held for approval do not silently spend the budget; they resolve when a human decides, and an approval releases a one-time token for that vendor and amount. Denied requests add nothing to the total but do add a receipt, which is what lets you distinguish a cap that stopped a runaway from a cap that merely annoyed a busy agent.

Sizing: normal day, honest worst day, hard stop

Start with the normal day's total from real prices: quantities multiplied by unit costs for a typical workload. Then model the honest worst day — the busiest legitimate day you can describe, with deadlines and volume spikes, but without malfunction. Set the cap between those numbers, usually nearer the worst honest day. The gap between the cap and the normal day is your malfunction budget: spend that only happens when something is wrong, and that the cap converts into denials and evidence instead of charges. If legitimate work repeatedly hits the cap, the receipts will show allowlisted vendors and coherent task references, and raising the cap is a considered decision rather than a guess.

Daily caps and human approvals working together

The approval threshold and the daily cap answer different questions. The threshold asks whether this amount deserves a human's attention. The daily cap asks whether today has any room left at all. A request can pass the threshold test — small enough to flow automatically — and still be denied because the day is spent. That ordering protects the approver as much as the budget: without it, a human can be socially engineered by a queue of individually reasonable approvals that sum to an unreasonable day. In MandateGuard the inbox shows pending items, while the engine keeps the day total impartially, so approving everything in front of you still cannot exceed the mandate.

Reading a day that hit the cap

When a day ends at the cap, read the chain in order. Early allowed purchases with coherent task references followed by denials suggest a genuinely heavy day that outgrew the budget. Repetition of the same vendor and amount suggests retries or a loop — fix the agent's error handling before touching the cap. A sudden appearance of new task references suggests scope creep: the agent took on work it was never mandated for. Each diagnosis has a different remedy, and only one of them is raising the cap. The hash-chained receipts make this review trustworthy, because the sequence you are reading can be verified as the sequence that happened.

Multiple agents, multiple days

On Pro and Team, each agent's policy carries its own daily cap, so a research agent's heavy day cannot consume the purchasing agent's headroom, and each day's story stays attributable to one agent and one job. Teams sometimes coordinate caps so their sum matches a departmental budget. Vendor trust and duplicate detection, both computed from your own ledger, get more useful as daily history accumulates: the same data that enforces today's cap becomes next quarter's evidence for how each vendor and agent actually behaved.

Five actions to take now

  1. Total a normal day's spend from real prices before setting anything
  2. Set the cap near the busiest honest day, not the average day
  3. Keep the per-task cap in place — daily caps do not stop single big mistakes
  4. When a day hits the cap, diagnose from receipts before raising it
  5. Give each agent its own daily cap when you run more than one agent

Worked example: the retry loop that cost 36 dollars, not 300

A data agent has a per-task cap of 50 dollars and a daily cap of 120 dollars. Its normal day is five batches at 16 dollars: 80 dollars. One afternoon its enrichment tool starts returning empty results, and the agent retries the 16-dollar batch every few minutes. The first seven retries are allowed — the day reaches 112 dollars of the 120-dollar cap — and the eighth request is denied because it would exceed the day.

The agent's operator sees the denial receipts stacking up with the same vendor, amount and task reference, recognises a loop rather than demand, and fixes the tool integration. Total damage: bounded, receipted and explained. Without the daily cap, the same loop running overnight at that pace would have attempted thousands of dollars of identical batches, each one individually legal.

Frequently asked questions

When does the daily cap reset?

With the new day. The historical chain does not reset — past days remain verifiable in the receipt ledger and in evidence exports on paid plans.

Do denied requests count against the daily cap?

No. Only allowed spend accumulates. Denials create receipts so the pressure is visible, but they do not consume budget.

Does an approved request count at request time or approval time?

It resolves through the approval flow: approval releases a one-time spend token for that vendor and amount, and the spend is recorded as the decision completes.

What if my agent works across time zones?

Pick the cap with your operating day in mind and review receipts by timestamp. The control that matters is the bounded total per day window, consistently applied.

Can the daily cap replace the per-task cap?

No. A daily cap of 200 dollars still permits one 200-dollar mistake. The two caps guard different shapes of loss and belong together.

Keep reading

Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides