Scaffold Heavy vs Model Trusting approach while building AI applications

When LLMs were new and less capable, to get something useful out of them required guiding the LLM towards the right path till the goal was achieved.

So at that time frameworks like Langchain got popular.

This approach is known as scaffolding heavy.

But as models became powerful enough, models didn't need the scaffolding to get something done. Instead the scaffolding came as a resistance in the general purpose applications. Hence giving minimal prompt and a tool list to the LLM was enough in most general cases.

This approach is known as model trusting.

Prime example of model trusting application is Claude Code.