MCP is great for integrating with Claude/Cursor, but building production agents with it does not make sense to me: You don't have access to the server's prompts, you lack observability, and can't debug.<p>Just had a convo with my teammate about this and curious what hn thinks.
Most of the work when trying to build a reliable agent is 1) determining which tools do you provide to the llm 2) how to describe the tools, their interface. MCP gives you pre-built tools that you can't change neither the interface nor the descriptions/prompts.<p>There is value in quick integration (or at least the promise of), but I don't see why it would be used when building an agent.