ExecuGraph: A Multi-Agent Framework for Execution-Grounded Backend Code Synthesis with Large Language Models

Read full text Summarise with VIRAT
Authors: Jothi Prabha Appadurai, Sai Deekshith Lekkala, G Suganeshwari
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Modern Large Language Models (LLMs) have shownstrong performance in generating backend code, but they often failto guarantee correctness, algorithmic efficiency, and runtimereliability due to reliance on single-agent, single-pass generationwithout validation. This paper introduces ExecuGraph, a multi-agent framework for execution-grounded backend code synthesis.The framework decomposes the generation into multiplespecialized agents organized within a directed workflow graphwith explicit state management. Correctness is enforced throughexecution-based evaluation, combining deterministic benchmarktests for known algorithmic problems with LLM-generated testcases for previously unseen tasks. ExecuGraph is implementedusing the LangGraph workflow orchestration framework andLangChain for model interaction, locally hosted LLM inferencevia the Ollama platform to ensure reproducibility, and optionallyintegrates a retrieval-augmented generation mechanism usingChromaDB for complex tasks to retrieve relevant algorithmicpatterns, while simpler problems are handled without databaseaccess. Experimental results on representative graph algorithms,dynamic programming, and data structure problems demonstratethat ExecuGraph significantly improves correctness, robustness,and interpretability compared to single-agent baselines,improving correctness from 76% to 93% in dynamicprogramming, and from 60% to 87% in graph algorithms, whilereducing execution failures in graph tasks from 66.7% to 0%.These results highlight the effectiveness of execution-groundedmulti-agent orchestration for robust AI-assisted backend codegeneration.

Fields of Study & Subjects

Software Engineering Research Teaching and Learning Programming Topic Modeling

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
22 March 2026
Language
EN