Effective LLM context length caps near 100k tokens across models despite 1M+ advertised windows

Garrit's post documents the split between a functional smart zone and degraded dumb zone in live agent sessions. Token consumption from file reads, debug traces, and test output pushes coding workflows past the threshold rapidly. Vendors continue publishing larger maximum lengths without corresponding gains in sustained performance. RULER results quantify the drop: multi-needle retrieval and aggregation tasks fall from 85% at 32k to below 40% at 128k on frontier models. Chroma's internal traces confirm gradual attention dilution rather than abrupt failure. Auto-compaction mechanisms introduced in Claude Code and similar agents summarize after degradation has already begun, propagating errors from the weakened window. Artifact-based handoff methods, as implemented in obra/superpowers and mattpocock/skills repositories, move state into external named documents before the smart zone boundary. This approach keeps each session under the measured 100k limit. Scaling laws for context length have not closed the gap between theoretical capacity and operational reliability.