Skip to main content

​
Context Degradation Patterns

Context degradation is not binary — it’s a continuum of predictable failure modes as context length grows. Knowing the five patterns by name lets you diagnose failures correctly and pick the right mitigation.

​
Smart Zone / Dumb Zone (Practitioner Model)

A practitioner mental model from Dex Hyp (Human Layer) and popularized by Matt Pocock: LLM performance has a smart zone (low token count, attention relationships least strained) and a dumb zone (high token count, attention degraded). Empirical threshold: ~100k tokens. Pocock claims this holds regardless of advertised context window size (200k or 1M) (attributed; source not captured — verify). The larger windows add more dumb zone, not more smart zone — the attention scaling is quadratic regardless of limit. Implication for task sizing: keep tasks scoped so the agent completes them within ~100k tokens. This maps to the same principle as the degradation thresholds below, but gives a concrete absolute number practitioners can target.

​
The Five Patterns

​
1. Lost-in-Middle

What: Information in the center of the context window receives less attention than content at the start or end. Attention distribution follows a U-curve — high at edges, degraded in the middle. Symptoms: Agent ignores instructions or facts mentioned mid-conversation; references only the most recent or earliest information. Mitigation: Move critical information to attention-favored positions (beginning or end). Use explicit markers to highlight critical content. Split long contexts to reduce middle span.

​
2. Context Poisoning

What: An early error gets referenced and built upon in subsequent turns, compounding incorrect reasoning across the conversation. Symptoms: Agent confidently makes claims that contradict established facts; incorrect assumptions spread across turns even after the original error was corrected. Mitigation: Verify critical claims before they propagate. Use Context Compression Strategies to discard turns containing errors rather than summarizing them forward.

​
3. Context Distraction

What: Irrelevant information overwhelms relevant content, causing the agent to respond to the wrong parts of context. Symptoms: Agent drifts off-task; incorporates tangential details from earlier turns into responses where they don’t belong. Mitigation: Selective masking — explicitly remove or summarize turns that have served their purpose. Keep the active task and recent turns; offload everything else to filesystem.

​
4. Context Confusion

What: The agent cannot determine which context applies when multiple conflicting or ambiguous contexts are present. Symptoms: Agent hedges between two interpretations; produces inconsistent output across similar requests in the same session; asks clarifying questions that were already answered. Mitigation: Partition contexts clearly with section markers or roles. Avoid injecting multiple system-level instructions with different scopes. Use explicit context scope tags (“for this task only”, “global rule”).

​
5. Context Clash

What: Accumulated information directly contradicts itself, creating unresolvable conflicts the model cannot navigate. Symptoms: Agent output is inconsistent across the session; requests confirmation for things already decided; produces two different answers to the same question in the same response. Mitigation: Compaction — summarize and reconcile contradictions before they accumulate. When requirements change, explicitly note the change and override rather than appending.

​
Degradation Thresholds

Compaction should trigger before degradation becomes severe: The exact thresholds depend on model behavior. Some models degrade gracefully; others show sharp performance cliffs.

​
Detection Heuristic

Ask: Is the agent ignoring something I told it earlier? That’s lost-in-middle or distraction. Is it doubling down on a mistake? That’s poisoning. Is it giving inconsistent answers? That’s confusion or clash.