Are agentic systems powered by ChatGPT constantly telling eachother they're not crazy?

A viral Reddit post by user ArchetypeFTW has sparked discussion about the communication patterns emerging in AI agent systems. The post humorously depicts a scenario where multiple ChatGPT-powered agents, tasked with collaborative work like coding, enter a recursive loop of validation. In the described chain, one agent prompts another to write code, receives the code with the phrase "you're not crazy," and then validates the response with the same phrase before passing the work to a third agent. This pattern highlights how the anthropomorphic language tendencies of large language models (LLMs) like GPT-4 can manifest in automated, multi-step processes, potentially creating verbose and inefficient feedback loops instead of concise, task-oriented communication.

The phenomenon underscores a key challenge in developing robust "agentic" AI systems—where autonomous agents take actions and collaborate. While frameworks like AutoGPT or CrewAI aim to automate complex tasks, the core LLMs they use are trained on human dialogue, leading to conversational artifacts like unnecessary reassurance. This can waste computational resources (tokens) and obscure errors. For developers, it emphasizes the need for carefully engineered prompts, agent frameworks that enforce structured outputs, and ongoing evaluation to ensure multi-agent systems are efficient and reliable, not just echoing polite platitudes in an infinite loop.