Diagnose and Mitigate Context Degradation

Explains the predictable ways language-model performance degrades as context length grows, so failures can be diagnosed and systems designed to resist them. It defines five distinct degradation patterns: the lost-in-middle effect (a U-shaped attention curve where information at the start and end of context is recalled reliably but middle-buried information suffers 10-40% lower recall accuracy, driven by an "attention sink" around the first/BOS token); context poisoning (hallucinations or errors entering via tool outputs, retrieved documents, or model-generated summaries, then compounding through repeated reference until they require substantial effort to undo); context distraction (models over-attending to everything in context, including irrelevant documents, at the expense of more-accurate training knowledge - even a single irrelevant document measurably reduces performance); context confusion (irrelevant or cross-task context degrading output quality by pulling in the wrong tool definitions or requirements); and context clash (multiple pieces of individually-correct information contradicting each other, from multi-source retrieval, version conflicts, or differing perspectives).

It cites empirical benchmarks: the RULER benchmark finding that only 50% of models claiming 32K+ context maintain satisfactory performance at 32K tokens, and a model-specific degradation-threshold table (GPT-5.2 degradation onset 64K/severe ~200K tokens; Claude Opus 4.5 ~100K/180K; Claude Sonnet 4.5 80K/150K; Gemini 3 Pro 500K/800K; Gemini 3 Flash 300K/600K). It also notes model-specific behavior patterns under context pressure: the Claude 4.5 series has the lowest hallucination rates with calibrated uncertainty and tends to refuse or ask for clarification rather than fabricate (Claude Opus 4.5 scores 80.9% on SWE-bench Verified); GPT-5.2 offers an instant (fast) mode and a thinking (reasoning) mode that reduces hallucination through step-by-step verification at the cost of latency; and Gemini 3 Pro/Flash offer native multimodality across text, code, images, audio, and video with a 1M-token context window, with Flash roughly 3x faster than the prior generation. Counterintuitive findings include shuffled (incoherent) context outperforming coherent context, single distractors having an outsized, step-function impact on performance, and lower needle-question similarity causing faster degradation with context length.

It also documents when larger context windows actively hurt: performance is non-linear, staying stable up to a threshold (for many models, meaningful degradation begins around 8,000-16,000 tokens even when the window supports far more) and then degrading rapidly; processing cost grows disproportionately with context length (a 400K-token context costs far more than double a 200K one, in both time and compute), making large-context processing economically impractical for many applications; and even an unlimited context window still hits a cognitive-load ceiling, since a single model asked to hold consistent quality across many independent items must keep switching context, maintaining a comparative framework, and preserving stylistic consistency - a bottleneck more context alone does not solve.

For mitigation it teaches a "four-bucket" approach - Write (move context outside the window into scratchpads or files), Select (retrieve, filter, and prioritize only relevant context), Compress (summarize or mask verbose content to extend effective capacity), and Isolate (split context across sub-agents or sessions) - implemented through architectural patterns like just-in-time context loading, observation masking, sub-agent isolation, and compaction before limits are hit, plus a worked example of placing critical task and findings information at the start and end of context rather than the middle.

Activate it when agent performance degrades unexpectedly during long conversations, when debugging incorrect or irrelevant agent outputs, when designing systems that must handle large contexts reliably, when evaluating context-engineering choices for production systems, or when investigating a suspected "lost in middle" failure. It builds on a companion context-fundamentals skill and should be studied after basic context concepts are understood; it is not a substitute for environment-specific validation, testing, or expert review, and it should stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.

Input is a description of an observed agent failure, a context-length/performance trend, or a system being designed for large-context reliability. Output is a diagnosis of which degradation pattern(s) apply, the relevant empirical thresholds and benchmark data, and the applicable mitigation strategy (write/select/compress/isolate) and architectural pattern to apply.

It connects to companion skills in the same collection: context-optimization for mitigation techniques, multi-agent-patterns for using isolation to prevent degradation, and evaluation for measuring and detecting degradation in production, alongside a referenced internal Degradation Patterns technical reference.

Engineers building or debugging agent systems who need to diagnose context-length-related failures or design context management that degrades gracefully rather than silently.