A New Research Thread: Agents for Dynamical Graph Grammars

This week I started a third research thread, and I learned a lesson about the order in which you write a paper.

Agentic DGG conversion

Dynamical Graph Grammars (DGGs), the stochastic modeling framework from Eric Mjolsness's group at UC Irvine, describe systems like growing microtubules or particle dynamics as graph rewrite rules. The same model can live in several representations: research math in LaTeX, simulation languages, and other tooling. I want to know whether a set of cooperating LLM agents can translate between them.

I laid out the architecture: an orchestrator that routes requests to specialist agents, a Pure Math agent for the LaTeX side, and RAG over the DGG literature so the agents are grounded in the actual formalism. I built the LLM client, a Chroma vector store, and PDF ingestion that scopes retrieval per agent. By midweek I had pivoted the first concrete target to FoxFlow, a simulation DSL for DGGs, and wired up an agent that sends requests to a FoxFlow transpiler service.

Everything is scaffolding at this point. The plan also lists more ambitious pieces, like learning from compiler feedback, which I'm deliberately leaving for later.

RAG saturation

I drafted the Results and Failure Analysis sections before the 30k-document run had finished. Some of the sentences I wrote, including a specific recall drop and a range of stability values, turned out not to match the data once it arrived. I spent a good part of the summer walking those back. The lesson: a draft results section should contain placeholders, not plausible-sounding numbers.

Lost in the Middle

I turned the experiment log into a draft reproduction report covering the setup, bucket definitions and results tables.