context-degradation

Context degradation occurs when long-running agents lose information quality or relevance over extended conversations—patterns like lost-in-middle, poisoning, and distraction reduce model performance and reliability. Detecting and mitigating these patterns maintains accuracy in production deployments.

Run degradation detectors on active sessions to identify patterns where critical information gets buried, contaminated, or obscured. The Skill diagnoses attention and data flow to pinpoint where context fails, enabling targeted remediation.

Context isolation, compaction, masking, and partitioning reduce degradation by restructuring how information flows through the agent. Adjust prompts and context windows based on detector findings to maintain reliability in long-running workloads.

Yes. Apply degradation detectors to production deployments, debugging scenarios, and large-context workloads to measure performance and generate actionable guidance for real-time prompt and window adjustments.

Larger windows amplify lost-in-middle effects and increase poisoning risk. Isolation and partitioning strategies help mitigate these constraints, but monitoring and compaction remain essential to maintain agent reliability at scale.