Prompts need to be tailored for specific model (duh). Tool definitions are de facto part of your system prompt. By injecting tool definition from MCP servers into your prompt, you are basically adding prompt that was (likely) not tailored for your specific model. That causes drop in quality of answers.

Other issue is that adding one new feature via MCP can introduce regression into your system and cause other MCP features to work incorrectly (and I am not talking about malicious prompt injection by MCP provider).

Real life example: tools from MCP server for ms outlook requires certain date format for filtering emails. That worked fine in our prod. Until we added another MCP server (built in-house in another department of our company) that required dates in different format. After adding these tools, now our agent started making mistakes: putting dates in outlook format in our internal mcp tool and vice versa - putting dates in our internal format into outlook tool. There is no obvious way to fix it other than separating these prompts. And that goes against standard MCP architecture and consumes way more tokens to cycle through multiple prompts and agents.

If that's already the issue with two MCP servers, then this architecture clearly does not scale.