ExecuGraph Framework Enhances Backend Code Synthesis with LLMs

While Large Language Models (LLMs) can generate plausible backend code, a single-pass approach offers no inherent guarantee of correctness or runtime reliability. This limitation necessitates a more robust framework for code synthesis. ExecuGraph is introduced as a multi-agent framework that places execution-based validation at the core of backend code synthesis. It employs six specialized agents—Planner, Code Generator, Logical Reviewer, Evaluator, Optimizer, and Explainer—coordinated through a typed directed workflow with a bounded retry budget. The framework is implemented using LangGraph with locally hosted models (Ollama) and can optionally include a retrieval layer for algorithmic technique recall. A subprocess-isolated sandbox with a wall-clock timeout ensures safe and controlled evaluation of generated code. Evaluations were conducted on a curated 30-problem DSA suite, HumanEval, and an APPS-introductory subset. While results on the internal DSA suite showed no statistically significant difference compared to single-agent baselines, ExecuGraph demonstrated a +3.1 percentage point improvement on HumanEval. The most compelling finding was cross-model: with DeepSeekCoder V2 Lite, the graph-category accuracy jumped from 57.5% (one-shot) to 80.0% (multi-full), a substantial +22.5 percentage point increase. This supports the hypothesis that the value of multi-agent decomposition scales with the base model's capability. The framework's primary contribution is its methodological design, allowing for controlled measurement of each component's marginal contribution through configurable one-shot, execution-retry, and per-agent ablation conditions.