Execution Grounded Multiagent Systems for Reliable Backend Code Generation with Large Language Models'

<title>Abstract</title> <p>Multi-agent architectures for large-language-model code synthesis are widely reported to improve correctness, but they are rarely compared against a single-agent baseline that receives the same execution feedback inside the same runtime. The gain attributed to role decomposition may therefore be the gain from execution feedback under another name. We present ExecuGraph, a workflow framework whose value here is as an instrument: one codebase collapses by configuration into single-agent one-shot (SO), single-agent execution-retry (SR), and full multi-agent (MF) conditions, so the two levers can be separated rather than confounded.On the complete 164-problem HumanEval benchmark, the only suite in our study with enough paired problems to be adequately powered, execution feedback is worth a large and unambiguous gain (SR over SO: +25.6 pp, 95% CI [+18.9, +32.3], exact Wilcoxon p &lt; 0.0001 after Holm adjustment), whereas adding five further agents on top of it is worth nothing measurable (MF over SR: −1.2 pp, CI [−7.3, +4.9], p = 0.76) at 3.6× the model calls. The second contrast is the more secure of the two because both conditions receive execution feedback and are therefore equally affected by the fact that the retry loop observes the scoring tests.A matched-compute control rules out the obvious alternative explanation: five independent one-shot samples filtered through the same sandbox, at five times the call budget, gain +0.4 pp, so the effect is feedback rather than extra sampling. We corroborate with a 30-problem algorithms suite, an APPS-introductory subset, and a second backbone, and we use three architecturally inert ablations as negative controls to measure run-to-run variance directly rather than assuming it. All figures are post-correction of a documented sandbox import-policy defect, re-scored from stored artifacts without re-invoking any model (Section 5.12). Configurations, per-trial logs, and table generators are released.</p>