Giving an agent more context can make it slower and more expensive without making its answer better. The key distinction here is between what a reader processes and what the model actually sees: a reader may examine a large source while exposing only the evidence needed for the current decision, with source anchors preserved.
Key Findings
- Across six models and five scenarios, token use fell by 15.7–92.9% and wall time by 23.0–89.0%; task scores were preserved or improved in 29 of 30 settings. The broad range matters: savings depend on the task and source, rather than coming from a single universal compression trick.
- On web deep-search tasks, token use dropped by 31.0–43.0% and completion time by 23.0–37.0%, while all six models gained 5–14 percentage points. This suggests selective reading can help even when agents must gather evidence across sources.
- The design separates reading control from evidence acquisition. A Read Intent Protocol tracks what the agent needs and what sources have covered; interchangeable readers acquire evidence; a budgeted gate limits cumulative exposure and falls back to a faithful full view when needed. This avoids treating retrieval or summarization as sufficient on its own.
- Results across three agent frameworks show lower token use and wall time in all 18 reported settings, with scores preserved or improved in 15. The approach is designed to integrate without retraining the model.
Who It Suits—and the Trade-off
A good fit for teams building tool-using agents that repeatedly inspect long pages, files, or structured data, especially when context and latency are bottlenecks. It is less compelling when an agent genuinely needs to inspect an entire source, when no suitable reader exists for the content type, or when a task depends on details that selective evidence acquisition might miss; in those cases the system falls back to a full view, and savings can shrink.