A large part of an AI agent's bill is not thinking but reading. When an agent opens a webpage, a PDF or a tool response, it usually swallows the whole thing, and that text then rides along in every later model call. Prior work cited by a new preprint puts external observations at 60% to 80% of the tokens agents process. SparseRead is a proposal to make agents read the way people do: decide what they need, look at the relevant passage, and open the full document only when that is not enough.
SparseRead sits between the agent and its sources and needs no change to the model. The agent states what it is looking for in a short read intent; the layer returns a bounded pack of evidence tied to its source, plus a record of what has and has not been inspected, and the agent decides whether that suffices. Different reader back ends handle text and PDFs, tables and JSON, multi-file collections and code. A budget gate caps how much raw content an episode can expose and falls back to the full view when partial reads would cost more, for example on small files.
The authors tested 30 settings, six models across five scenarios, including Claude Opus 5, GLM-5.1 and Kimi-K2.5. Tokens fell by 15.7% to 92.9% and wall time by 23.0% to 89.0%, with medians of 61.4% and 53.7%. Quality held or improved in 29 of the 30 settings; the exception was GLM-5.1 on a science task. The biggest savings came on long documents. On web deep search, the most familiar workload, tokens fell by 31.0% to 43.0% and time by 23.0% to 37.0%, while scores rose by 5 to 14 points, a sign that less clutter can mean better answers. Against retrieval and context-trimming baselines, SparseRead was the only method that kept or improved scores in every setting tested.
For anyone paying for agents through Claude Code, Codex or an API, this targets real money, and the finding that even a strong model like Opus 5 saved 41.0% to 89.0% of tokens suggests better models do not make the problem go away. It also fits a wider pattern of agent infrastructure moving from bigger context windows to smarter control of what goes into them.
The caveats are the usual ones for a preprint. The scenarios and the agent framework, NanoBot, were chosen by the authors, and the headline 92.9% comes from long-document reading, the case most favourable to the idea. Real web pages are messier than benchmark tasks, and a reader that misses the one paragraph that mattered fails quietly. The work has not been peer reviewed.