Skip to content

Latest commit

 

History

History
192 lines (131 loc) · 4.72 KB

multi-agent-code-review.md

File metadata and controls

192 lines (131 loc) · 4.72 KB

Multi-Agent Code Review

Pattern: Launch parallel agent reviews before presenting code to users.

Problem: Single-perspective code review misses issues. Serial reviews are slow. Users receive code with obvious problems that could have been caught.

Solution: Launch multiple specialized agents in parallel. Each focuses on a different quality dimension. Collect findings, fix critical issues automatically, then present production-ready code.

How It Works

flowchart LR
    A[Code] --> B[Parallel Review]
    B --> C[Collect Findings]
    C -->|Critical| D[Auto-Fix]
    D --> B
    C -->|Clean| E[Present]
Loading

Key Characteristics

Aspect Description
Parallel execution All agents run simultaneously
Specialized focus Each agent reviews one dimension
Auto-fix loop Critical issues fixed before presentation
Confidence scoring Only high-confidence issues flagged

Agent Specializations

Architecture Advisor

Reviews structural decisions:

  • Component boundaries and coupling
  • Dependency direction (no circular deps)
  • Pattern consistency across codebase
  • API design and contracts

Simplification Advisor

Reviews for unnecessary complexity:

  • Over-abstraction (premature generalization)
  • Dead code and unused exports
  • Redundant logic and duplication
  • Opportunities to simplify

Security Advisor

Reviews security posture:

  • Input validation at boundaries
  • Authentication/authorization checks
  • Secret handling (no hardcoded credentials)
  • OWASP Top 10 vulnerabilities

Implementation

Reference Implementation

A reference orchestrator implementation would include:

  • Parallel agent dispatch
  • Finding collection and aggregation
  • Critical issue auto-fix loop
  • Report generation

Note: Adapt the agent invocation syntax to your specific tooling (Claude Code Task tool, OpenAI Assistants API, etc.). See the demo-fastapi repository for working examples.

Agent Prompt Template

You are the [SPECIALIZATION] Advisor.

Review the following code changes for [FOCUS AREA] issues.

## Rules
1. Only flag issues you're confident about (>80% certainty)
2. Provide specific file and line references
3. Categorize findings as CRITICAL, WARNING, or INFO
4. Include fix suggestions for CRITICAL issues

## Code to Review
[CODE_DIFF]

## Output Format
### CRITICAL
- [file:line] Description. Fix: suggestion

### WARNING
- [file:line] Description

### INFO
- [file:line] Observation

Finding Categories

Category Action Threshold
CRITICAL Auto-fix before presenting Block presentation
WARNING Include in review notes User decides
INFO Optional mention Low priority

Auto-Fix Rules

Only auto-fix when:

  1. Issue is CRITICAL (blocks presentation)
  2. Fix is mechanical (not judgment call)
  3. Confidence is high (>90%)
  4. Fix can be validated

Examples of auto-fixable:

  • Missing input validation
  • Unused import removal
  • Hardcoded secret detection (replace with env var)

Examples requiring human judgment:

  • Architecture refactoring
  • API design changes
  • Performance trade-offs

Measuring Success

Key Metrics

Metric Before Target How to Measure
Critical issues in presented code 2-3 0 Agent review findings
Rework cycles per feature 3-5 1 Iteration count
Time spent on code review Hours Minutes Review duration

Success Indicators

  • User never sees code with critical issues
  • First code presentation is production-ready
  • Reviews complete in <10 minutes (parallel agents)

Integration with AQE

Multi-agent review integrates with Autonomous Quality Enforcement:

flowchart LR
    A[Code] --> B[Validate]
    B --> C[Review]
    C -->|Issues| D[Fix]
    D --> B
    C -->|Clean| E[Commit]
Loading

Sequence:

  1. Validation loop ensures code compiles and tests pass
  2. Multi-agent review checks quality dimensions
  3. Critical issues feed back into validation loop
  4. Only fully validated and reviewed code reaches user

Benefits

Benefit Description
Parallel speed 3 reviews in time of 1
Specialized depth Each agent focuses on one area
Consistent standards Same review criteria every time
Auto-remediation Critical issues fixed automatically
Clean presentations Users see production-ready code

Related Patterns