TLDR: A recent ZDNET report highlights that even advanced AI agents are struggling to effectively utilize the Model Context Protocol (MCP), a middleware designed to integrate AI models with external applications. This challenge leads to performance delays and underscores the need for more specialized training for AI models to navigate complex tasks via MCP.
Leading artificial intelligence agents are encountering substantial difficulties when interacting with the Model Context Protocol (MCP), an increasingly prevalent middleware designed to bridge AI models with various external applications. A ZDNET report, published on October 14, 2025, reveals that despite MCP’s intent to enhance generative AI programs like chatbots by enabling connections to resources such as databases and customer relationship management software, even the most sophisticated AI models are struggling with its implementation.
Introduced last year by generative AI startup Anthropic (creators of the Claude family of large language models), MCP was envisioned as a secure, industry-standard method for LLMs and AI agents to interface with external software. As ZDNET’s Steven Vaughan-Nichols explains, this middleware aims to streamline operations by reducing the number of individual connections an AI program needs to initiate for multiple external resources.
However, multiple studies cited in the report indicate that top AI models, including Google’s Gemini 5, require numerous rounds of interaction with external programs when operating through MCP. This iterative process results in significant delays in the AI models’ performance, hindering their efficiency and responsiveness.
Also Read:
- Google Unveils Data Commons MCP Server, Revolutionizing AI Access to Public Data
- Google Unveils Data Commons MCP Server to Bolster AI with Verified Public Data
ZDNET’s key takeaways from the findings emphasize several critical points: even the best AI models are challenged to carry out tasks effectively via MCP, new benchmarks demonstrate that models struggle particularly as tasks become more complex, and there is a clear requirement for more specific training of AI models tailored to MCP use. This suggests that while the protocol holds promise for expanding AI capabilities, current models lack the inherent understanding or training to fully leverage its potential, creating a bottleneck in the advancement of integrated AI agent functionality.
