Arbiter gives 152 results for the system prompts of 3 coding agents
Prompt architecture goes together with the failure class, but not with the severity, the authors write.
Claimed, not confirmed
Researchers made Arbiter, a framework that tests the system prompts of AI coding agents for interference patterns. They used it on Claude Code from Anthropic, Codex CLI from OpenAI, and Gemini CLI from Google. An undirected scan gave 152 results. In a directed analysis of one vendor, the researchers found 21 interference patterns and labeled them manually. The authors write that the total cost was $0.27.
Sources
Posted
