A Multi-Agent Framework for Evidence-Grounded Automated Code Review
Reviewers trust a comment when they can see where it came from. The paper puts a group of LLM agents behind that rule: nothing reaches the review until it can be traced to retrieved evidence in the code.
→ reference implementationcitation
Hakan Bogan, Ibrahim Arikan, Emrah Inan, Onur Demirors. “A Multi-Agent Framework for Evidence-Grounded Automated Code Review.” ICENSTED 2026 Proceedings Book, 3rd International Conference on Engineering, Natural Sciences, and Technological Developments (ICENSTED 2026), Erzurum, Türkiye, 2026, pp. 137-146. ISBN 978-625-7960-90-8.
- affiliation
- Department of Computer Engineering, İzmir Institute of Technology, İzmir, Türkiye
- conference
- 3rd International Conference on Engineering, Natural Sciences, and Technological Developments
- held
- 3–6 July 2026, Erzurum, Türkiye · organised by Atatürk University
- isbn
- 978-625-7960-90-8
- pages
- 137-146
- keywords
- Multi-agent systems · Large language models · CrewAI · Automated code review · Static analysis · Software engineering · Evidence-based reasoning
abstract
Large language models are increasingly integrated into software engineering workflows, but their use in automated code review remains limited by hallucination, lack of evidence grounding, and inconsistent reasoning across complex changes. This study proposes a CrewAI-based multi-agent code review framework composed of specialised agents — Change & Context Analyst, Security Reviewer, Style & Format Reviewer, Revision Proposer, and Supervisor. The system enforces a strict evidence-based review paradigm in which every finding must be supported by static analysis output (Semgrep, Bandit, Ruff, ESLint), file-line references, and relevant code snippets. Evaluated across diverse pull request scenarios using precision, recall, F1-score, and review efficiency, the architecture produces actionable and consistent review output while maintaining strong precision and structured reasoning.