← Go back

Guide // MandateGuard

How do you give a research agent an API budget that works?

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 research agent's budget works when it is led by the daily cap and indexed by question. Fit the per-task cap to the largest routine data pull, the daily cap to the busiest honest research day, and reserve approvals for the two events that deserve humans: bulk purchases and first use of a new source. Require task references naming the research question, so receipts join spend to shipped output. The monthly spend-versus-output review — cutting sources that feed nothing, defunding repeated questions — is where the budget earns its keep, because research waste is quiet redundancy, not dramatic overspend.

What does a research agent spend on?

A research agent's budget is its MandateGuard policy fitted to research economics: small, frequent API purchases governed mainly by the daily cap, with approvals at bulk-purchase boundaries and first-use of new sources. Research value is judged separately — by whether the output was used — and the receipt history, organised by task reference, is what joins spend to output for that judgement.

Research spend: many small doors

Research agents buy breadth: search calls, page fetches, data API queries, model runs, the occasional dataset. Each purchase is small enough to feel free, which is exactly the risk — nobody approves a 4-dollar call, and four hundred of them are a budget. The control centre of gravity therefore shifts from the per-task cap (rarely binding) to the daily cap and the ledger. The daily cap bounds the agent's curiosity per day; the ledger, organised by task reference, makes curiosity accountable per question. Set those two properly and research spending becomes legible: you can price a question, compare providers per question, and notice when the agent re-asks questions the organisation already paid to answer.

Fit the numbers to the work's rhythm

Derive the daily cap from the busiest honest research day: a deep-dive day with its searches, fetches and model runs priced out. Set the per-task cap at the largest routine purchase (a full data pull), and the approval threshold where bulk begins — the dataset buy, the archive access, the premium tier trial. Expect the threshold to be crossed rarely and meaningfully; if bulk buys are weekly, they are part of the work's rhythm and the threshold is misplaced, or the work wants a standing arrangement reviewed as procurement. Keep the allowlist to the providers your researchers actually trust for citable data; research agents are champion source-collectors, and every source they can pay for unsupervised is a source whose quality you have implicitly endorsed.

Task references are the research budget's index

For research, the task reference should name the question or the ticket the research serves. That one discipline converts the receipt chain into a research accounts system: this question cost 64 dollars across three providers; this report's evidence cost 121 dollars; this dataset was bought twice for two questions (legitimately) or twice for one question (a denial that should have happened). Without references, research spend is a single undifferentiated pool and every budget conversation is impressionistic. With them, the monthly review can do the most valuable thing a research budget review can do: join spend to output and defund the questions and sources that produce nothing.

Approvals for novelty and bulk, silence for routine

Research dies under heavy approval friction — the whole point is fast, broad inquiry — so the approval design must be surgical. Routine calls flow. Two events pause: bulk purchases, because they concentrate spend in one judgement; and first use of a new source, because source quality is a standing endorsement disguised as a transaction. Both are written-condition territory in the approver runbook (and on envelopes for defined research projects). Approvers decide bulk buys with the question's context — has this been answered, is this source the right one — which is a thirty-second judgement when the task reference does its job. Everything else about research supervision is retrospective: the monthly review, not the daily queue.

The spend-versus-output review

Monthly, group receipts by task reference and lay them beside the research artefacts actually shipped: reports, briefs, decisions informed. Three findings recur. Sources with spend and no shipped output — cut or re-scope them. Questions researched repeatedly — create the internal note that makes the answer findable, and let approvers deny the third asking. Providers whose data is always superseded — the market has spoken; adjust the allowlist. This review routinely pays for the entire mandate programme in this domain, because research waste is not dramatic overspend but quiet redundancy, visible only when spend is indexed by question. The receipts provide the index; the review provides the will.

Project research and envelopes

Defined research projects — the market study, the due-diligence sprint — deserve envelopes for the same reasons client projects do: the budget belongs to the project, the sources may exceed the standing allowlist, and someone will ask what the study cost. The envelope carries the total, the period and the approval conditions (bulk buys, new sources), and closes into a case file that sits beside the report it funded. Standing research continues under the policy. Teams that adopt this split find their research conversations change altitude: standing spend is tuned for efficiency, project spend is judged for value, and neither contaminates the other's numbers.

Five actions to take now

  1. Fit caps to API price bands: routine calls flow, bulk jobs approve
  2. Set the daily cap from token-and-tool spend on the busiest research day
  3. Require task references naming the research question or ticket
  4. Approve first use of any new data source as a written condition
  5. Review receipts against research output monthly; cut sources that produce nothing

Worked example: the provider whose data never shipped

A product team's research agent runs with a per-task cap of 40 dollars, a daily cap of 110 dollars, approve-over at 25 dollars and four data and model providers allowlisted. A normal week produces about 90 receipts, mostly 3-to-15-dollar calls. Two bulk dataset buys (32 and 38 dollars) pause for approval; the lead approves one and denies the other after seeing the same ground covered in a previous report.

The monthly join is the payoff: receipts grouped by task reference against the reports actually shipped. One provider's 210 dollars of calls fed no shipped report — its sources were always superseded. The provider leaves the allowlist; the agent's daily cap drops to 90 dollars on the evidence; and the research budget starts the next month 15 percent cheaper with no output lost.

Frequently asked questions

Why is the daily cap the main control for research?

Because research spend is many small purchases. Per-task caps rarely bind; the day's aggregate is where money accumulates.

How do I stop the agent re-researching the same question?

Task references naming the question make repeats visible in receipts; approvers deny the bulk buy that duplicates a shipped report, as in the example.

Should the agent have browsing or premium sources beyond the allowlist?

Only through approval. First use of a new source is a novelty decision a human should make once, deliberately.

What if research needs a 200-dollar dataset?

That's an envelope or a one-off human-led purchase — a considered acquisition, not a bigger standing cap.

How often should sources be pruned?

Monthly, at the spend-versus-output review. A source with spend and no shipped output for a month is a candidate for removal.

Keep reading

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