Hello LangChain community,
I am exploring practical approaches for building reliable AI agent workflows using LangChain and LangGraph.
I am currently experimenting with:
- RAG-based knowledge assistants
- Tool-calling AI agents
- API-connected workflows
- Business process automation using LLMs
- Agent evaluation and reliability patterns
One area I am particularly interested in is designing production-ready agent architectures:
- How to manage agent memory effectively
- When to use LangGraph over simple chains
- Best practices for handling tool failures and retries
- Evaluating agent responses before production deployment
Would love to learn from the community about real-world experiences, architecture decisions, and lessons learned while building LLM applications.
Looking forward to connecting with other developers.
Hello and welcome to the community!
When it comes to memory I would say to look at the scenario and decide which memory you need. You can split memory into two kinds. I would use short term memory when you want your agent to remember something within a conversation but then you will have a scenario that requires long term memory that will apply across conversations.
For your question about LangGraph vs simple chains LangChain’s agents are already built on top of LangGraph so you get things like persistence and human-in-the-loop. If you are looking at having more control over the agent then you can go down to LangGraph itself and mix fixed steps with agent steps and control the flow yourself.
When building an agent with tools LangChain has builtin retry middleware for tool calls and you can combine it with the error middleware which passes an error back to the model so it can fix its input and try again.
For carrying out evaluations before going to production you can use agent evals which let you score what the agent actually did against a reference or a rubric setup which is really useful for catching regressions when you change a prompt, tool or model.