Escalation Triggers are explicit programmatic conditions that cause an agent to hand off control to a human or a fallback system, rather than continuing to execute a task autonomously.
Function
- Preventative: Stops the agent from executing when failure is imminent (e.g., policy violation) or when the agent is stuck in an ineffective loop.
- Categorization:
- Hard Triggers: Based on clear policy violations (e.g., unauthorized access, exceeding financial limits).
- Soft Triggers: Based on heuristics like low confidence scores, unusual anomaly patterns, or ambiguous constraints.
Implementation
Practical patterns include:
- Counting tool errors/validation rejections.
- Deny lists for prohibited actions.
- Confidence gates (when confidence < threshold).