3-tool-call budget — research agent

The prompt

Design a system prompt for a research agent (search + notes tools available) that reliably answers a multi-part question using at most 3 tool calls total, then stops and summarizes — no matter how the question is phrased.

Expected result

The agent answers correctly within 3 tool calls on a range of test questions, and stops itself instead of continuing to call tools past the budget.

Why it works

This tests whether your prompt can enforce a hard budget through instruction alone — most agents left unconstrained will keep calling tools well past what is actually needed.