Thank you for such an in-depth exploration of the MCP ecosystem! It's fascinating to see how MCP is changing the landscape for AI agents beyond just chat interactions. The comparison to Coca-Cola's journey with infrastructure really highlights the importance of synchronized development between applications and infrastructure.
I wonder if the point about MCP solving the impedance mismatch between LLM's probabilistic output and structured need of apis is the key benefit as mentioned in the example "without MCP" and "with MCP".
That problem is solved by tool calling capability of LLM which is induced during instruction tuning while training LLMs. MCP based agents eventually also uses this capability to decide which MCP service (tool) to call and what parameter values to provide.
Rather MCP helps loose couple the actual tool implementations from the use, making MxN problem into M+N. Once a tool/service is wrapped into an MCP server, anyone can use it in a standard way instead of coding the whole thing up himself/herself.
Great thought piece! If you are interested in collaborating on building MCPs, reach out to me at jonathan@jeantechnologies. We are building MCP tools for businesses as well.
Love this post. Disruptive enablement is what protocols do.
John, thanks so much! Looking forward to catching up soon.
Thank you for such an in-depth exploration of the MCP ecosystem! It's fascinating to see how MCP is changing the landscape for AI agents beyond just chat interactions. The comparison to Coca-Cola's journey with infrastructure really highlights the importance of synchronized development between applications and infrastructure.
@haley thank you so much for reading! It’s indeed amazing how much history rhymes, from Coca Cola through to ChatGPT
This is the best explanation of the MCP developments that I've seen, even for the non-technical crowd.
Nice explanation.
I wonder if the point about MCP solving the impedance mismatch between LLM's probabilistic output and structured need of apis is the key benefit as mentioned in the example "without MCP" and "with MCP".
That problem is solved by tool calling capability of LLM which is induced during instruction tuning while training LLMs. MCP based agents eventually also uses this capability to decide which MCP service (tool) to call and what parameter values to provide.
Rather MCP helps loose couple the actual tool implementations from the use, making MxN problem into M+N. Once a tool/service is wrapped into an MCP server, anyone can use it in a standard way instead of coding the whole thing up himself/herself.
Great thought piece! If you are interested in collaborating on building MCPs, reach out to me at jonathan@jeantechnologies. We are building MCP tools for businesses as well.