Every token in the window competes with every other. Include what is load-bearing for the current step; rank and prune the rest.
First recorded July 17, 2026 · Last revised July 17, 2026
Rationale
Context windows grew a thousandfold and the instinct became: include everything, let the model sort it out. But attention is not free — irrelevant context dilutes relevant context, buries the instruction that mattered, and raises cost and latency on every call.
The emerging rule: treat context as a budget to be allocated, not a bin to be filled. Every document, every tool result, every historical turn should earn its tokens by being plausibly load-bearing for the current step. Retrieval should rank and prune, not dump.
This doctrine is still forming. Open questions include how aggressively to summarize versus quote, and where the break-even sits between one bloated call and several focused ones.