Funny how this article from Anthropic "Code execution with MCP: Building more efficient agents" says that MCP - which they invented - doesn't work so well...
Some of their examples use Salesforce. And they're sending record updates inside loops. 🤦♂️ 😂
MCP was like the "low code" theme for LLMs.
From the outside, many people appeared to see MCP that way: a universal interface with tools an AI could "just call" with natural-language orchestration and no need for custom glue code.
Same vibe of the no-code/low-code, "resolve everything in Salesforce with flows" and "no devs".
But MCP tends to consume a lot of tokens redundantly and is slow - Anthropic itself now admits.
The solution they propose now is to have LLMs generate code that runs agents in an environment where tools are only loaded when needed and the data flow is coordinated and processed locally before passing it back to a LLM.
That brought to my mind 2 articles I read: "Reality has a surprising amount of detail" and "The truth is always made of details".
You can't abstract details away - not all the time, not always neatly.
MCP would give LLMs a clean interface to tools, but it can't hide the real system underneath and the need for auth, schema, rate limits, handling of execution errors and side effects, I/O boundaries, sandbox constraints, state management, etc. - each of them will come to the surface inevitably/eventually.
