The article compares working with large language models (LLMs) to collaborating with human coworkers, emphasizing that both can misinterpret vague instructions. It discusses the importance of clear communication and proper context when interacting with LLMs, suggesting that many frustrations stem from unrealistic expectations of deterministic behavior. Adapting to this probabilistic nature can lead to more effective outcomes.

MCP (Model Context Protocol) has gained significant attention as a standard for LLMs to interact with the world, but the author criticizes its implementation for lacking mature engineering practices, poor documentation, and questionable design choices. The article argues that the transport methods, particularly HTTP and SSE, are problematic and suggests that a more straightforward approach using WebSockets would be preferable.