Any reason for writing everything from scratch and not using Langgraph?

[adsk2050]

Langgraph and Langchain make it very easy to build agents. Any specific reason for writing everything from scratch?

For example, OpenAI's Structured outputs API practically ensures the output schema of the data and langchain has good support for it. To write everything needed to make it work from scratch seems a bit cumbersome and redundant, doesn't it?

[omkar-334]

langchain and llamaindex offer too much abstraction. When you want to debug, it becomes cumbersome. Writing it from scratch gives more control and we can debug every step of the process

[Vishnu8299]

Hi @adsk2050,

That’s a great question — and you’re right that tools like LangGraph and LangChain do simplify building agents, especially with features like structured outputs and integrated memory/chaining.

However, writing everything from scratch was a deliberate choice for a few key reasons:

🔍 Why from scratch?
Full control & transparency – When building from the ground up, we understand exactly how each step behaves, which makes debugging and optimization much easier.

Avoiding unnecessary abstraction – While LangChain and LangGraph are powerful, they often introduce layers that can obscure what’s actually happening under the hood.

Custom logic & flexibility – We needed to handle very specific use cases and edge conditions that would’ve been difficult to support cleanly with existing abstractions.

Learning & research value – Building agents from first principles helped us learn more deeply and tailor the solution to our architecture and performance goals.

That said, for many projects — especially MVPs or smaller experiments — LangChain/LangGraph can be incredibly productive. It really comes down to the balance between speed of development and depth of control.

Thanks again for the thoughtful question!

Best,
vishnu vardhan