Agentic Code Review: A Multi-Agent Framework with Meta-Cognitive Reflection and Human-in-the-Loop Alignment

Automated code review remains a persistent challenge in software engineering. Traditional static analysis tools depend on rigid pattern matching and produce excessive false positives, while large language model (LLM)-based approaches frequently hallucinate findings that reference nonexistent code. This paper presents a multi-agent framework that addresses both limitations through orchestrated specialization and adversarial self-verification. The proposed system deploys four agents via LangGraph: a security specialist that combines Bandit static analysis with LLM-based contextual reasoning, a retrievalaugmented performance optimizer that grounds suggestions in historical pull request data, a meta-cognitive critic that validates all findings against the original source code to filter hallucinations, and a lead architect that synthesizes only verified results. A human-in-the-loop checkpoint ensures that critical decisions remain under expert oversight before final synthesis. Evaluation on deliberately vulnerable Python code samples demonstrates that the critic agent reduces false positives by 40 % relative to a single-agent baseline, improves line-number accuracy from 67 % to$\mathbf{9 2 \%}$, and lowers the hallucination rate from$\mathbf{3 2 \%}$to$\mathbf{1 8 \%}$. An ablation study confirms that the meta-cognitive verification layer contributes most significantly to these gains. These results indicate that adversarial self-verification within multi-agent architectures can meaningfully improve the reliability of LLMdriven code analysis for high-stakes software review tasks.