Prefer Code Over MCP Tools

Model Context Protocol tools are not truly composable and require excessive context, making them less efficient than direct code execution.

Writing code directly often yields faster, more reliable automation than relying on LLM inference for each step.

Automating tasks via generated scripts allows for easier validation, debugging, and repeated execution at scale.

A practical pipeline uses an LLM to generate code, run it, and then have an LLM review results, ensuring correctness.

Future improvements may involve better abstractions combining code generation with LLM judgment for non-developers.