GSoC 2026 Proposal: Mesa-LLM Stabilization (175hr) - Feedback Request

[savirpatil]

Hi @colinfrisch and @jackiekazil (and any others who would like to give feedback on my proposal),
I've been working on my GSoC proposal for the Mesa-LLM stabilization project and would really appreciate some early feedback before I finalize it. Here's a summary of my planned approach for the 175-hour scope:

Background: I've built three Mesa/Mesa-LLM models from scratch to understand the library and some current friction points. These include a pollinator-plant cascade extinction model (PR #352), a misinformation spread model with LLM vs rule-based comparison (PR #384), and a basketball strategy simulation (in progress). Building these surfaced several issues as well as giving me an overall better understanding of the Mesa code I would be working with.

Phase 1: Outstanding Issues and Mesa 4.x Compatibility - 4 weeks (~50 hours)
Address the most impactful blocking bugs in the repo, focusing on Mesa 4.x compatibility breaks that prevent the library from running on the latest version of Mesa. Each fix incudes a regression test added to the existing CI testing.

Phase : LLM Tool Modularization and Local Inference 3 weeks (~50 hours, Midterm)
Improve the modularity and reliability of the core reasoning and memory abstractions, making structured output more accessible and local inference support more straightforward for new users. Focus on reducing the gap between what the library promises and what users can actually accomplish without workarounds.

Phase 3: Documentation and Scientific Examples 4 weeks (~55 hours)
Build out comprehensive documentation including a rewritten getting started tutorial, a migration guide from plain Mesa to Mesa-LLM, and at least one example model sourced from a scientific paper that demonstrates the full Mesa-LLM stack. Add performance and cost optimization guidelines for researchers running simulations at scale.

Phase 4: CI and Stable Release Cadence 1 week (~20 hours)
Expand test coverage, add CI matrix testing across Mesa versions, and establish a documented release process aligned with Mesa's cadence so users can depend on version compatibility going forward.

General Questions:
For the reasoning and memory improvements in Phase 2, is there a preferred architectural direction you have in mind, or is it open for proposal?
For the scientific paper example in Phase 3, do you have a paper or domain in mind that would best showcase Mesa-LLM's capabilities?
Are there open issues you'd prioritize differently for the 175-hour scope?

Thanks for your time and I would be happy to share the full proposal draft if helpful!

[jackiekazil]

We had gotten many requests for feedback and conversations, and it is impossible for us to provide it because of time constraints because we would selecting at random, who we gave feedback too and it wouldn't be fair.

Proposals do not have to perfect and if accepted, we will work with you on a more detailed execution path and schedule.

[savirpatil]

Thanks for the clarification and and it’s good to know that proposals don’t need to be perfect since the details can be refined later if they’re accepted!