My name is Raaghav, and I am a software engineer and an ex-founder.
I am interested in making a contribution to LangGraph. Before I started coding away I wanted to get an idea of if my effort would be wasted.
I’ve been studying the recent Recursive Language Models (RLM) work on inference-time scaling for arbitrarily long prompts, and I believe there’s an opportunity to introduce an opt-in “environment-mediated recursive executor” within LangGraph that aligns with its graph-based orchestration model.
The idea would not be to modify default agent behavior, but to provide a new executor pattern that:
• Treats large prompts or tool outputs as an external environment (chunked + indexed)
• Allows the model to iteratively retrieve, decompose, and solve over bounded snippets
• Enforces hard constraints (max depth, max tool calls, token/cost budgets, timeouts)
• Provides deterministic termination and full execution traces
• Benchmarks against flat-context and naive retrieval agents
LangGraph already supports cyclical graphs and recursion limits, which makes it a natural substrate for this. What’s missing today is a standardized, budget-aware recursive execution pattern that scales beyond model context windows in a principled way.
I would propose building this as:
A new opt-in executor/graph template (e.g., RecursiveEnvironmentExecutor)
A structured memory layer (chunk graph with provenance)
A BudgetManager enforcing cost and depth constraints
A benchmark suite demonstrating improvements on long-log and tool-heavy workflows
Would this contribution be something useful? Im happy to start coding away. I wanted to get some feedback first
This sounds like a very strong and aligned contribution, especially since LangGraph already supports cyclic graphs and recursion control but doesn’t have a standard, budget-aware recursive executor pattern. I would suggest starting a design discussion issue first to make sure the scope is correct with the maintainers, but this direction is definitely useful and not a waste of time.
Hi @Bitcot_Kaushal just to confirm I should open a new issue on GitHub? I believe I have to be a privileged user to do this, correct? It says I need to confirm I am a LangGraph maintainer, or was asked directly by a LangGraph maintainer to create an issue here. Could you walk me through the official process to create a design discussion issue ?
Hey @raaghavchaudhry1 — on the GitHub procedural question, the “must be a maintainer” gate you hit is almost certainly tied to a specific template (likely the RFC / Design Discussion one — those are commonly locked to maintainers in LangChain repos). The repo doesn’t have a Discussions tab enabled either, so the open path for non-maintainers is to open a regular Feature Request issue using the default template, link this forum thread in the body, and explicitly frame it as “design discussion needed before implementation.” Tag a maintainer or two for visibility.
Separately — I’ve been studying RLM from a different angle and have been building production multi-agent pipelines using LangGraph for the past few months as the technical co-founder in stealth (especially long-context fanout and retrieval + synthesis loops). Your proposal overlaps with several pain points I’ve hit directly, and I’d genuinely like to be a co-contributor on this if you’re open to it.
Would be glad to contribute to this development. And if you’re up, we could sync on scope and discuss more — happy to do that async over GitHub once the issue is up, or hop on a quick call, whichever helps. Let me know how you’d like to coordinate.