"Not Worth Another Token": Pruning Deep-Research Agent Context Before Retrieval Cuts Token Usage Up to 73% With Little Quality Loss

arXiv (via HuggingFace Daily Papers)·medium signalXBlueskyLinkedInCopy link

An arXiv paper (2608.08389, submitted August 9) from a team including Adobe Research authors tests marginal-value estimation for pruning at three points in a deep-research agent pipeline — pre-retrieval, post-retrieval, and pre-synthesis — comparing lightweight heuristics against a learned value model. Lightweight heuristics alone cut token usage up to 73% with little quality degradation, and early-stage pruning produced by far the largest end-to-end savings; later-stage pruning mostly just tightens the synthesis context. No single strategy won across quality, efficiency and faithfulness at once, which is the honest caveat. For anyone running multi-agent research pipelines on a token budget, the actionable result is that the cheapest win is deciding what *not* to retrieve.