How Multi‑Agent AI Transforms Code Review into Automated Fixes

In modern software development, code review is essential for quality assurance and team collaboration, yet it often suffers from fragmented information, low manual efficiency, and insufficient automation. Lint results, tests, issues, and change logs are scattered across different systems, making holistic analysis difficult.

Unified Review Engine Integrating Lint, Tests, Issues, and Structural Analysis

AutoDev introduces an Agentic architecture that consolidates these disparate sources, enabling a closed‑loop workflow from analysis to automatic fixing.

Four‑Step Review Pipeline

Static Information Collection (Code Audit) : Extracts changed hunks from git diff, locates affected classes/methods with CodeGraph, runs ESLint/Ktlint/Detekt, and aggregates issues, requirements, and test data into structured output without consuming LLM tokens.

AI‑Driven Analysis (Code Analysis) : Constructs a system prompt from diff, structure, lint, issue, and test data; performs multi‑dimensional analysis (comprehensive, performance, security, style); supplements context via readfile or grep when needed; outputs structured ReviewFinding objects with severity, location, and suggestions.

Modification Plan Generation : Extracts key problems and repair paths, merges lint and AI findings, ranks them, and uses a dedicated prompt to produce an executable plan (to‑do list, steps, priorities) that users can refine before applying.

Automatic Fix Generation : Based on actual changed hunks, generates patches; aggregates plan, lint, analysis, and user feedback; the CodeReviewAgent coordinates the CodingAgent to perform multi‑round tool calls (read/write files, refactor, test) and produces final patches with metadata, supporting rollback and iterative refinement.

Multi‑Agent Collaboration for Platform Engineering

The system consists of a primary CodeReviewAgent that orchestrates the overall task and several specialized sub‑agents:

SubAgents (Analysis Agents)

AnalysisAgent – handles large content blocks and complex context to reduce token pressure on the main agent.

ErrorRecoveryAgent – recovers from tool failures or model errors.

CodebaseInvestigatorAgent – scans the codebase, performs cross‑file location and structural extraction to provide a global view.

CodingAgent (Repair Agent)

Executes read/write, lint, test, and other tool calls to apply code modifications.

Generates patches based on analysis output and returns structured repair results.

Sub‑agents are created and destroyed by SubAgentManager, while all tools are registered in ToolRegistry and dispatched by ToolOrchestrator, which also handles permissions and result processing.

Typical Workflow

The main agent receives a review request and aggregates diff, lint, issue, and test information.

If the content is large, it delegates to AnalysisAgent for preprocessing.

The main agent builds a prompt and invokes the LLM for analysis.

During the repair phase, the task is handed to CodingAgent to modify the code.

If errors occur, ErrorRecoveryAgent attempts to recover.

Information Aggregation Foundations

AutoDev automatically collects and merges multiple sources of data:

Git Diff – precise identification of all changes in a commit or PR.

Static Analysis (Lint) – runs various linters and ranks results by severity.

Issue Tracking – links related issues and extracts descriptions, statuses, and labels.

Test Relevance – analyzes the relationship between changed code and test files to assess risk.

Code Structure Analysis – uses tools like CodeGraph to identify affected functions, classes, and other structural units.

All gathered information is fed into the LLM prompt in a structured format, ensuring the model sees the full context for professional judgment.

Future Outlook

AutoDev aims to further integrate test coverage, CI/CD pipelines, incremental analysis, and other capabilities, continuously pushing the boundaries of intelligent development tools and positioning AI as a “super reviewer” for development teams.

Installation command (CLI): npm install -g @autodev/cli then run autodev review -p. Download the AutoDev Desktop from the GitHub release page (compose‑0.3.1).