Why I Chose LangGraph and LangChain Over Other AI Agent Frameworks for Lab Work?

A breakdown of the frameworks I evaluated for our multi-agent AI work and why LangGraph and LangChain fit best for my research.

When setting up our Agentic Lab, one of the first things I had to decide was which framework to use for building AI agents. There were a couple of options, but in the end, it boiled down to LangGraph for research and AutoGen for user-facing workflows.

TL;DR

After testing multiple frameworks, I chose LangGraph and LangChain for building and researching AI agents, while I plan to use AutoGen for agentic workflows where I’m the user.

  • LangGraph is the right level for building agent logic.
  • It allows deep customization of agent behavior.
  • Great selection of available models.
  • Plenty of pre-built tools for common workflows.
  • Custom tool development is straightforward.
  • LangGraph Studio and LangSmith provide useful debugging and tracking tools.
  • AutoGen is a strong option for agentic workflows where I’m interacting with the agents rather than building them.

LangChain has the most stars on GitHub (at least for now), while AutoGen Studio looks promising for cases where I need a more user-friendly approach. I’ll definitely experiment with it in the future.

What I Tested and Why They Didn’t Fit

I evaluated several frameworks, making my shortlist with ChatGPT before running actual experiments.

  • LangGraph – Probably the most widely used structured AI workflow framework. Built on top of LangChain, it allows for fine control over agent interactions, supports graph-based execution, and comes with LangGraph Studio, which makes it easy to visually track and debug agent behavior.

  • CrewAI – A simpler orchestration framework for multi-agent setups. It’s easy to use and great for quick deployments, but I found it too opinionated for my needs. It abstracts away too much, making it harder to tweak agent behaviors or experiment with non-standard workflows.

  • ParlAI – A research-oriented framework from Facebook AI that looked promising in terms of vision, but it was archived in 2023, which made it a no-go.

  • AutoGen – Developed by Microsoft, AutoGen looks extremely promising. The no-code builder is especially tempting, as I’ve been using ComfyUI for a long time. However, I didn’t choose it (for now) because:

    • The learning curve felt steeper compared to LangGraph.
    • Tooling seems less pre-built, requiring more setup.
    • LangGraph’s workflow fits my thought process better.
      That said, I’ll definitely revisit AutoGen, especially its Studio feature.

  • Custom Python Code – I considered building my own framework from scratch, which would have given me full control. However, I quickly realized this would mean spending too much time maintaining low-level implementations instead of focusing on actual research.

Why I Chose LangGraph + LangChain

After testing, LangGraph and LangChain stood out as the best fit for my research work. The key reasons:

  • Flexibility & Customization – LangGraph allows me to define complex agent interactions, making it easy to simulate multi-agent behavior with structured workflows.
  • Integrated Debugging & Visualization – The Studio feature in LangGraph is a so practical for tracking agent behavior over time, which is crucial for studying emergent interactions.
  • Active Ecosystem & Community Support – LangChain has a well-maintained library, plenty of integrations, and an active user base. This means I can iterate faster rather than reinventing basic components.
  • Pre-built tooling – LangGraph/LangChain come with ready-to-use tools, speeding up development so I can focus on core agent interactions.
  • Engineering feel – LangGraph feels like it’s built by engineers for engineers. The logic aligns with how I think, and building agents feels natural. While I like node-based interfaces (like ComfyUI), LangGraph is simply the best tool to get the lab up and running fast.

What I’ll Use It For

This setup will power multiple research directions at the Agentic Lab:

  • Multi-Agent Simulations – My first project is developing multi-agent conversation simulations for behavior modeling, starting with baby-daddy interactions. LangGraph makes it easy to define personality shifts and feedback loops over time. It’s also straightforward to run large-scale interaction cycles, log results in a structured way, and build datasets for behavior and personality development research, reports, and visualization.

  • AI Agent Evaluation – I’m building a framework to evaluate AI agent traits, starting with personality trait evaluation for workspace safety simulations. This involves structured dialogues, measuring behavioral patterns, and integrating public datasets. LangGraph makes data handling and evaluation integration seamless.

  • Research Automation – I’ve built a research agent to assist with web and academic research. LangChain’s integrations make it easy to expand its capabilities, adding PDF analysis and refining output control.

Final Thoughts

LangGraph and LangChain offer the right balance between structured execution and open-ended experimentation, making them the best fit for my work. I didn’t want a black-box system or an overly rigid framework—I needed something that let me test, iterate, and refine without fighting the tools themselves.

After building my first few agents, I realized LangGraph is the right tool for experiments where I need custom logic, data collection, or evaluation tools.

On the other hand, for workflows where I am the user, not the builder, I see AutoGen Studio being a better fit. My research automation agent, for example, might be simpler and more user-friendly in AutoGen Studio. I’ll likely transition to it in the coming weeks for those use cases.