Agent4cs: Multi-agent code summarization, up to 38% gains

Agent4cs, a multi-agent framework for code summarization submitted to arXiv on 1 July 2026, tackles large hierarchical codebases by delegating tasks to specialized agents rather than using a single model. The system was evaluated on seven frontier models and accepted to the main track of the 23rd European Conference on Multi-Agent Systems (EUMAS 2026).

What is Agent4cs and how does it work?

Agent4cs is a bottom-up multi-agent summarization framework that splits responsibilities across three agents: a summarization agent, a keyword-extraction agent, and a quality-assurance agent. The summarization agent focuses on producing robust summaries, the keyword-extraction agent proactively identifies critical information from subfolders, and the quality-assurance agent iteratively refines outputs for readability, coherence, and completeness.

The paper positions Agent4cs against common single-model approaches, noting existing solutions often rely on a single language model or coding assistant like Claude Code and treat source code as flat text, which underuses repository interdependencies and hierarchical structure. Agent4cs instead leverages those hierarchical relationships to assemble summaries from folder-level units upward.

How does Agent4cs perform compared with structured prompting baselines?

Agent4cs improves semantic consistency across all folder levels by an average 8% compared to two structured prompting baselines that use code segments, and shows up to 38% gains in normalized keyword coverage rate over the same baselines. The authors evaluated the framework on seven frontier models and report these improvements across folder levels and on real-world datasets.

Those two performance figures are the paper's key quantitative claims: an average 8% semantic-consistency gain across folder levels, and up to 38% improvement in normalized keyword coverage rate. The evaluation is presented as both multi-model (seven frontier models) and dataset-driven (real-world datasets), which the authors use to demonstrate consistency and keyword coverage improvements relative to structured prompting baselines.

Why it matters

Agent4cs tackles two common failure modes in code summarization: flattening hierarchical context and relying on a single assistant. By explicitly extracting folder-level keywords and iteratively checking summary quality, the framework aims to preserve repository interdependencies and surface higher-importance terms. For teams facing large, poorly documented or obfuscated codebases, improvements in semantic consistency and keyword coverage could make generated summaries more trustworthy and actionable.

What to watch

Look for the conference presentation and the full paper at EUMAS 2026, where the approach and evaluation details will be available; the arXiv submission is arXiv:2607.01425 (submitted 1 Jul 2026). Also check how the seven frontier models were configured in the authors' experiments and whether the authors publish code or data linked from the paper.

Authors and provenance: the paper is authored by Yongjian Tang, Ezgi Sarikayak, Doruk Tuncel, Jie M. Zhang, and Thomas Runkler, and was submitted to arXiv on 1 July 2026. The work was accepted to the main track of the 23rd European Conference on Multi-Agent Systems (EUMAS 2026).