Arbiter: Detecting Interference in LLM Agent System Prompts
Arbiter is a framework that combines formal rules and multi-model LLM analysis to detect interference patterns in AI agent system prompts. Analyzing three major coding agent prompts revealed that prompt architecture dictates failure modes: monolithic prompts cause growth-level bugs, while modular ones cause design-level bugs. The research shows that system prompt engineering requires software-like infrastructure for consistency.