As someone deep in the LangChain ecosystem, I’m wondering if Microsoft AutoGen still relevant after LangGraph for my specific use case. LangGraph seems like a natural extension of LCEL, but I really like how AutoGen handles code execution. Is it worth maintaining two different libraries in one project for these specific strengths?
3 answers
For LangChain power users, LangGraph is definitely more "idiomatic," but Microsoft AutoGen still relevant after LangGraph because of its unique built-in tools like the code executor. While you can replicate code execution in LangChain, AutoGen makes it a first-class citizen. Agents can write, debug, and run code in a sandbox natively. If your agent's primary task is data analysis or technical automation, AutoGen’s multi-agent collaboration around a code-executor node is often more robust and easier to set up than building equivalent custom tools in a LangGraph node.
Pamela, have you experienced any compatibility issues when trying to wrap AutoGen agents inside a LangGraph node? I'm trying to get the best of both worlds!
AutoGen still wins on code-heavy tasks. LangGraph wins on workflow-heavy tasks. It really depends on if your agent is a "thinker" or a "doer."
Well put. I use AutoGen for my data science agents because the code-execution loop is just so much more refined out of the box.
Joshua, it is possible but messy. You essentially treat the AutoGen group chat as a single "tool" within a LangGraph node. It works, but state sharing between the two is a headache. I’d recommend picking one for the core logic and only nesting them if you have a very specific legacy requirement.