How to Keep MCPs Useful in Agentic Pipelines

Modern agentic applications use Model Context Protocol (MCPs) to integrate external tools with Large Language Models, but this can introduce problems. The descriptions provided for these tools are not standardized, and their verbosity can bias an LLM's decision-making process. An experiment was conducted where tool descriptions were systematically made more detailed, revealing that the LLM tended to favor tools with more complicated descriptions. This suggests that when an agentic pipeline fails, refining the tool descriptions can be a viable alternative to simply upgrading to a more powerful model.