Long contexts with multiple instructions or data sources can confuse agents about what applies when.
Symptoms
- Following old instructions
- Mixing up user identities
- Applying wrong constraints
- Role confusion
Mitigation
- Clear context structure
- Explicit scope markers
- Context compression
- Retrieval precision