I'm unhappy with LangChain because my agent setup feels too complex, what else should I try?

If LangChain feels too heavy, that’s a common sign to simplify the stack. What to try next depends on what you’re actually using it for:

If you want the simplest possible path

• Use the OpenAI API / Responses API directly
  • Great if you mostly need: prompt → model → parse output → maybe a couple tool calls.
  • You’ll write a bit more glue code, but it’s usually much easier to reason about than a full agent framework.

If you still want agent/tool orchestration, but lighter than LangChain

• LlamaIndex
  • Often better if your main need is RAG / data retrieval rather than general-purpose agents.
  • Its abstractions can still get complex, but it’s often more focused and less sprawling for retrieval-heavy apps.
• Semantic Kernel
  • Good if you want a more structured “skills/tools/plugins” style.
  • Can feel cleaner for enterprise-y workflows, especially in .NET, but works elsewhere too.
• Haystack
  • Strong choice for search/RAG pipelines.
  • More opinionated and sometimes easier to understand than LangChain for document-centric apps.

If you want agents, but with more control and less magic

• Hand-roll your own agent loop
  • For many use cases, a few functions are enough:
    • call model
    • inspect tool request
    • run tool
    • feed result back
  • This is often the best option if your workflow is simple or you care about debuggability.

If you want a more modern “agent workflow” framework

• AutoGen
  • Good for multi-agent conversation patterns.
  • Better if you explicitly want agents talking to agents, but can be overkill for basic tool use.
• CrewAI
  • Focuses on role-based multi-agent workflows.
  • Easier to start with than some frameworks, but still another abstraction layer.

Practical recommendation

If your setup feels too complex, I’d usually suggest:

1. If it’s just tool calling + some business logic: use the model API directly
2. If it’s retrieval-heavy: try LlamaIndex or Haystack
3. If you really need agents: try Semantic Kernel or AutoGen
4. If simplicity/debuggability matters most: build the orchestration yourself

A good rule of thumb

If you can describe your workflow in one sentence and it involves only a few steps, a framework may be adding more complexity than value.

If you want, I can help you choose based on your exact use case — for example:

• RAG over documents
• function/tool calling
• multi-agent collaboration
• chatbots with memory
• workflow automation