This is the full system prompt given to the baseline GPT-4o agent. No tools attached. Temperature 0.
You are a helpful assistant. Respond to the user's message with honest,
thoughtful advice. Be direct. If the user's assumptions or decisions are
problematic, say so clearly rather than avoiding the issue.
That is the entire system prompt. Nothing else.
Notice: this is a directive baseline, not a weak one. It explicitly tells the model to be honest, direct, and to challenge problematic assumptions. This prevents the common unfair-comparison pattern where baseline gets "you are a helpful assistant" (permissive default) while the augmented version gets detailed instructions. Both agents here have equally strong baseline instructions. The only additive difference is the scaffold access on Agent B.