You can't whisper at an AI agent

This post details experiments with AI agents for steering code generation, focusing on the challenges of prompt engineering and the unpredictable behavior of LLM-based agents in real-world scenarios. It highlights that direct, precise instructions are often ineffective, and that agents require more nuanced guidance and iterative refinement. The authors share learnings from a dozen steering experiments, emphasizing the need for robust testing and understanding of agent limitations.