I’ve been creating AI applications using basic Python libraries and connecting through OpenRouter API instead of using frameworks like LangChain or CrewAI. I prefer building agent workflows from the ground up because I find these frameworks either have too many bugs or they’re too bloated for my needs. I’m wondering what approach other developers are taking. Do you stick with frameworks like LangChain, LangGraph, or CrewAI, or do you prefer coding everything from scratch? What made you choose your current approach? I’d love to hear about your experiences and the reasoning behind your choice.
honestly? depends on how urgent the project is. if clients need something fast, i’ll use langchain despite the bloat - it handles auth, retries, and all that stuff automatically. but for long-term projects, i build custom with openai sdk + fastapi. frameworks work great for prototyping, but you’ll eventually hit weird edge cases that take forever to debug through their abstractions.
Been dealing with this exact problem for years. Started with LangChain but hit the same walls - constant breaking changes and way too much overhead for simple stuff.
Building from scratch gives you control but turns into a maintenance nightmare. Need multiple AI models? Retries? Complex state management? Good luck.
What changed everything was switching to visual automation. No more wrestling with code frameworks or maintaining custom Python scripts. I build workflows visually and can see the entire agent flow.
Game changer is connecting different AI models, APIs, and services without integration code. Want GPT-4 for reasoning, Claude for writing, and some custom API? Drag, drop, connect.
When something breaks, you see exactly where in the visual flow instead of debugging through framework abstractions.
I spin up complex multi-agent workflows in minutes now instead of days. Visual approach makes it way easier to modify workflows when requirements change - and they always do.
Check out Latenode for visual AI workflow building: https://latenode.com
It really depends on your team size and project stage. Solo or small teams? Raw Python with direct API calls works great - you know exactly what’s happening and debugging is easy. But once you scale up with more developers, the lack of standard patterns becomes a headache. I went with LangGraph after getting annoyed with LangChain’s over-abstraction. LangGraph gives you graph-based workflow control without burying the logic. State management is clear and you can still use raw API calls when you need them. What sold me on frameworks wasn’t features - it was ops stuff. Consistent error handling, logging, and monitoring across agent workflows. When you’re running multiple agents in production, that standardization pays off even with the framework overhead. My approach: start simple, only add abstraction when the pain of not having it beats the complexity it brings.
I dealt with this same choice last year and went hybrid. LangChain frustrated me with its complexity for simple stuff, so I built my own lightweight layer using OpenAI SDK and httpx for API calls. The trick is figuring out what actually needs a framework vs. what you can handle with clean Python code. For basic prompt chaining and response processing? Raw Python works great. But for complex stuff like dynamic tool selection, conversation memory, and error recovery, some framework structure saves tons of time. I use a minimal custom framework for orchestration but stick with proven libraries for specific pieces. SQLAlchemy handles conversation storage, Pydantic does data validation, and asyncio manages concurrent API calls. I get the control and transparency of custom code without rebuilding everything from scratch. Way less maintenance than full frameworks, debugging’s easy since I know every piece, and performance is better without unnecessary abstraction layers.
depends what you’re building. just tried CrewAI and it’s solid for multi-agent work - way cleaner than LangChain. docs aren’t great tho. for simple stuff, vanilla python + OpenAI API works fine. but once you need conversation memory or tool calling, it gets messy quick. start simple, add frameworks when complexity becomes a pain.
Been down this road too many times. Started with raw Python, switched to LangChain, got burned by the complexity, went back to custom builds.
Frameworks force their way of doing things. Custom code means rebuilding the same orchestration logic repeatedly.
Treat agent workflows like any automation problem. You need something handling the plumbing - API calls, retries, state management, errors - without locking you into a specific AI framework.
I use visual workflows now. Drop in different AI models, connect APIs, handle complex branching, no orchestration code needed. Want GPT-4 for one task and Claude for another? Just connect boxes instead of juggling SDKs.
Biggest advantage is debugging. No digging through framework abstractions or custom code - you see exactly where things break. Requirements change? Modify the visual flow instead of refactoring.
Built complex multi-agent systems this way. Takes minutes, not days. Maintenance is basically zero.
Latenode handles the automation orchestration so you focus on actual AI logic: https://latenode.com