Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{198445,
        author = {Rohan Mistry and Ruchika Dungarani},
        title = {SafeCode-MAS: A Safety-Aware Multi-Agent Framework for Automated Software Development with Adaptive Memory and Iterative Refinement},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {11},
        pages = {13237-13246},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=198445},
        abstract = {The growing integration of Large Language Models (LLMs) into software development workflows has created an urgent need for intelligent systems capable of managing multi-step code generation while simultaneously ensuring output quality and security. Existing multi-agent frameworks such as MetaGPT, ChatDev, and AutoGen have demonstrated the value of role-based specialization for automated development, yet they uniformly lack integrated safety verification mechanisms, exhibit sequential tool chaining failure rates of 60–70%, and operate without persistent memory for cross-session learning. This paper presents SafeCode-MAS, a novel multi-agent framework comprising four specialized agents — Planner, Coder, Reviewer, and Safety — coordinated through structured communication protocols with bounded iterative refinement of maximum three cycles. The framework introduces a dual-layer memory system combining session-specific Short-Term Memory with persistent Long-Term Memory implemented through the ChromaDB vector database, and a dedicated Safety Agent integrating Bandit and Semgrep static analysis for Common Weakness Enumeration vulnerability detection. SafeCode-MAS is evaluated on HumanEval (164 tasks), MBPP (500 tasks), and Custom-30, a purpose-built benchmark of 30 multi-file development scenarios. Against four baselines with five-seed repetition and Wilcoxon signed-rank testing, SafeCode-MAS achieves 92.07% Pass@1 on HumanEval, reduces vulnerability density by 73.8% from 13.17 to 3.44 vulnerabilities per thousand lines, and demonstrates progressive learning with first-iteration quality scores improving from 6.8 to 8.4 across sequential domain tasks. These results establish safety-integrated multi-agent architectures as a data-driven approach to managing the quality, security, and efficiency of automated software development processes.},
        keywords = {Multi-Agent Systems; Code Safety; Software Development Automation; Large Language Models; Data-Driven Quality Management},
        month = {April},
        }