LLM performance degrades as input context grows — non-uniformly and even on simple tasks. The same question answered correctly at 1k tokens of context gets missed at 100k, and the degradation is not a smooth tail: distractors, position, and semantic similarity of the surrounding text all modulate it. "The context window is supported" ≠ "the context window is used well."
Implication for agent harnesses: a long-running main agent's judgment quality is a resource that depletes as its context fills — so context hygiene (eviction, isolation, summarization) is not bookkeeping, it is capability preservation. This is the load-bearing fact behind certificate-is-the-subagent-boundary: what crosses back from a subagent must be small, both to be checkable and to keep the parent's context from rotting.
Source: Chroma Research technical report 'Context Rot' (2025) — needle-in-haystack-style evals showing non-uniform degradation with input length; referenced from certificate-is-the-subagent-boundary during the 2026-07 wiki promotion
The starting-context side of this discipline: in a large codebase, what goes in at the start is the map, not the code — an index shaped like the directory tree plus bidirectional links between files, the two shapes code actually has. The agent gets the navigation operations an IDE gives a human (go to definition, find references) as tools, and a file enters the context only when the agent navigates to it. Everything else stays out, deliberately — preloading is how the rot starts. See give-the-agent-a-map.