Code Review Agents

Code Review Agents is the reference implementation of the paper "A Multi-Agent Framework for Evidence-Grounded Automated Code Review", published in the ICENSTED 2026 proceedings. It replaces the single monolithic reviewer most LLM tools ship with a team of specialised agents that each hold one concern.

A Change & Context Analyst reads the diff against the stated intent. A Security Reviewer combines Semgrep and Bandit output with model reasoning. A Style & Format Reviewer consolidates Ruff and ESLint. A Revision Proposer writes the suggested change. A Supervisor then deduplicates across agents, resolves conflicts, calibrates severity and filters noise before anything is shown.

The rule that holds the whole thing together is the evidence policy: a finding may not reach the review unless it points at static-analysis output or at an explicit file and line with the code attached. That constraint is what makes the output checkable rather than merely plausible, and it is the part reviewers actually trust.