context-degradation

Detect context poisoning in agent systems using poisoning detection tooling that analyzes conversation inputs for malicious or distracting content. This identifies injected distractions and confusion patterns to preserve operational reliability during production agent runs.

Analyze attention distribution in long-running agent sessions using Python scripts that measure focus across the context window. This provides degradation scores and actionable recommendations to monitor and improve operational performance.

You need numpy installed to run the provided degradation detector and analyze sample contexts. The Python scripts utilize this dependency to execute attention distribution analysis and calculate degradation scores for agent systems.

Use degradation scoring for debugging agent systems during long conversations and production agent runs. Apply the diagnostic tooling when context grows large enough to cause lost-in-middle, poisoning, distraction, or confusion patterns that impact reliability.