I like to view MCP as a higher lever of abstraction on tool usage. First came langchain with its tool functions, and now we have MCP servers that encapsulate a full feature and usability of a specific framework.
At first I thought that MCP were more related with the 'reasoning' part of the agent (well they maybe are bc the way tools are defined on the MCP server --markdown like instructions for each tool and the purpose of the MCP server-- they improve the overall performance of the agent).
But the most interesting fact about this is that Agent workflows will became more complex. Then we will improve even more the reasoning part (overall intelligence of the agents), maybe we will see more adversarial style of agents (just like MoE but on a higher level).
Interesting times ahead