Ext · Cross-cutting

State Corruption

Detects when agent memory or state becomes corrupted, including type drift, schema violations, nullification, and velocity anomalies. Second-highest accuracy detector.

Examples

  • Numeric field "price" suddenly contains a string value after processing
  • Critical state field value changes to None/null mid-workflow
  • Three or more tracked state fields disappear simultaneously
  • Field value changes direction 5 times in rapid succession (velocity anomaly)

Detection methods

Schema Validation
Checks state values against expected types and domain bounds
Nested Flattening
Recursively flattens nested structures (e.g., n8n json wrappers)
Velocity Analysis
Detects abnormal rate of state changes
Cross-Field Validation
Ensures relationships between related fields are consistent

Calibration accuracy

F1
0.906
Precision
0.955
Recall
0.863

From the Pisama calibration set. See detector scoreboard for the full table.

Detect this in production with the framework adapters (LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Claude Agent SDK, n8n, Dify). See the full taxonomy at /taxonomy.