MESA-LLM

[Spartan-71]

Hello everyone! I am interested in working on the mesa-llm project.
I have created this new discussion so that all the comments/discussions are in one place.

[Spartan-71]

As mentioned in the Project wiki mesa-llm should have reusable components like planning, memory and reasoning modules.

Can we use Langchain, Crewai, or any other agentic framework to integrate LLM into Mesa, or should we just refer to how these frameworks are built and then build relevant components from scratch?

[wang-boyu]

At this point we're open to ideas on whether to use third party frameworks or not

[WingchunSiu]

I was thinking the same thing. I think existing frameworks like LangChain and CrewAI are useful for structured workflows but may be too rigid for Mesa-LLM’s modular needs. They encapsulate the planning, memory, and reasoning a bit too much, which simplifies development but may limits deep customization for abm.

A more flexible alternative I could think of is LangGraph, which allows graph-based decision flows. For example adding an LLM-based supervisor to oversee agent reasoning can help maintain adaptability, but ultimately we need highly customizable components tailored to Mesa’s ecosystem. So rather than forcing an external framework it makes more sense to design modular and extensible agent structures I guess.

Curious to hear thoughts—do we see areas where existing frameworks would help, or is customization unavoidable for integration?

[colinfrisch]

Hey, I've worked with the HuggingFace's smolagent framework and it could be a reasonable start. Since the library is very easy of use, it could really simplify the coding compared to doing some function calling using sthg like the OpenAI API directly. The other advantage is reasoning (pre-implemented). In my opinion, it would not be incompatible with a LangGraph based structure. What do you think ?

from mesa import Agent

class LLMAgent(Agent):
    def __init__(self, unique_id, model, smol_agent):
        super().__init__(unique_id, model)
        self.smol_agent = smol_agent

    def step(self):
        task = "Describe the current environment state and --take action with available function tools--"
        response = self.smol_agent.run(task)


class SimulationModel(Model):
    def __init__(self):
        self.smol_agent = CodeAgent(tools=[TOOL_LIST], model=HfApiModel(model_id="MODEL_TO_USE"))

        for i in range(number_of_agents):
            agent = LLMAgent(i, self, self.smol_agent)
            self.schedule.add(agent)  

[piratevector]

@colinfrisch I think this snippet is using Mesa 2.x

[colinfrisch]

@colinfrisch I think this snippet is using Mesa 2.x

It's not a code that's currently implemented in mesa, just a suggestion for a feature :)

[Bhavishya-Chaturvedi]

Hello @Spartan-71 and @WingchunSiu, I'd like to collaborate on this project as well. I haven't found much discussion consolidated in one place. Are there any recent developments or a central discussion area I might have missed?

[WingchunSiu]

Hey, you can join Mesa's group discussion on https://summerofcode.withgoogle.com/programs/2025/organizations/project-mesa through the "chat" on the right.

[shubhambhar007]

Hi @WingchunSiu is there any ongoing discussion in github?
I am very excited to apply for this project in this upcoming GSoC and would really appreciate your guidance on the same.
Thanks!

[wang-boyu]

This is a good place to discuss Mesa-LLM. And of couse using matrix chat is also encouraged!

[Bhavishya-Chaturvedi]

Thanks a lot @shubhambhar007 @wang-boyu

[wang-boyu]

I have answeres some questions raised in a separate discussion about mesa-llm. Here is the link in case it's of interests to some others: #2046 (reply in thread)

[Bhavana-bhuvana]

Hi everyone,

I'm Bhavana S final year B.C.A student new to open source and really excited to start contributing. I’m currently exploring the MESA-LLM: Generative Agent-Based Modeling with LLM-Empowered Agents project for GSoC 2025 and preparing my proposal. I’ve started going through the Mesa documentation, some practical examples, and also the CONTRIBUTING.md file to understand how to get started the right way.

I have experience with Python, data analysis, and machine learning using Scikit-learn. I’m also exploring deep learning and LLMs, and have worked on frontend projects using Streamlit. Recently, I started learning React to expand my frontend skills.

I would really appreciate any guidance on how to study the Mesa framework and this project more effectively. If there are specific resources, issues, or tasks you'd recommend, I’d love to work on them. I’m flexible and open to learning, and can align my study path with your suggestions.

Thanks a lot, and I’m looking forward to contributing and learning more with the community!

[mawa-maria]

Hi @Spartan-71, and the Mesa team,

I'm Fatima-ezzahra, an AI/Computer Engineering student at ENSAM Casablanca, excited to apply for GSoC 2025 to work on Mesa-LLM. Here's how I plan to contribute:

Next Steps

1. First Contributions (starting this week):

   • Tackle "good first issues" to understand Mesa's architecture
   • Improve documentation (e.g., add LLM integration examples)

2. Technical Approach (for the project):

   class LLMAgent(Agent):
       def __init__(self, unique_id, model, llm_backend="huggingface"):
           super().__init__(unique_id, model)
           self.llm = HuggingFacePipeline()  # Default backend
           self.memory = ShortTermMemory()  # Customizable module

Focus on modularity (swap LLM backends easily)

Build debug tools (LLM decision logging)

Why I'm a Good Fit
Skills: Python, ML (IBM Data Science cert), basic Mesa knowledge

Projects: Built AI systems like [Calorie Tracker] (recommendation engine)

Availability: 56h/week dedicated to GSoC (May-Aug)

Questions
Which LLM backends are highest priority? (HuggingFace/OpenAI)

Are there example models I should study first?

I'll begin working through open issues and share progress here. Looking forward to your guidance!

Best,
Fatima-ezzahra
GitHub: mawa-maria

[Spartan-71]

hey @mawa-maria, great to know that you are also interested in MESA-LLM project.

Which LLM backends are highest priority? (HuggingFace/OpenAI)

As mentioned in the Project Wiki, I think there is no such priority from the Mesa team for the backend; they need integrations with as much as LLMs as possible. To ensure the same, we can use LiteLLM, which is a LLM Gateway / Proxy Server.

Are there example models I should study first?

I recommend you go through the introductory tutorial, which will help you in understanding "The Boltzmann Wealth Model" in detail, with the implementation.

[AlexandreFln]

Hi everyone, if it can be of any interest I tried to design an energy market simulation using LLM-powered agents with Mesa framework. This is a good start but it can largely be improved and I would be happy to hear your thoughts on it.

Here is the repo

[wang-boyu]

Thanks for sharing! We'll definitely look into it while working on mesa-llm in the following months.