Claude Multi-Agent Collaboration: Architect + Engineer + Reviewer + Optimizer Four-Role Pattern

When you type “help me build a system” into Claude, the model needs to simultaneously play the roles of architecture design, code writing, quality review, and performance optimization. The community has found that splitting these four roles into a pipeline significantly improves output quality.

Four-Role Division

This workflow splits a Claude conversation into four stages, each with independent role settings and output requirements:

Role	Responsibility	Output
Architect	Understand requirements, design system architecture, determine tech stack	Architecture diagram, module division, interface definitions, tech decision records
Engineer	Implement code based on architecture, write tests	Complete code, unit tests, integration tests
Reviewer	Review code quality, security, maintainability	Code review report, issue list, fix suggestions
Optimizer	Performance tuning, refactoring, documentation	Optimized code, performance benchmarks, usage docs

Why It Works

The core advantage is role isolation reducing cognitive load. When Claude focuses on a single role:

• The Architect doesn’t get bogged down in implementation details and can think more strategically about system design
• The Engineer, with clear architecture documents, implements more efficiently
• The Reviewer, in a “find faults” mindset, catches issues more easily than when the same model “self-reviews”
• The Optimizer works on stable code for performance improvements, avoiding the rework caused by “optimizing while writing”

Step-by-Step Guide

Step 1: Architect Phase

You are a senior system architect. Please design the system architecture based on the following requirements:

Requirements: [describe your project needs]

Please output:
1. System architecture diagram (describe module relationships in text)
2. Technology choices and reasoning
3. Module division and interface definitions
4. Potential risks and mitigation plans

Note: Output design only, no code.

Step 2: Engineer Phase

You are a senior engineer. Here is the system architecture design:

[Paste Architect's output]

Please implement the code based on the architecture design. Requirements:
1. Implement module by module
2. Each module includes unit tests
3. Follow the tech choices defined in the architecture
4. Code comments should be clear

Note: Strictly follow the architecture design. If there are issues, annotate them in code comments.

Step 3: Reviewer Phase

You are a senior code reviewer. Here is the implemented code:

[Paste Engineer's output]

Please review from these dimensions:
1. Code quality and readability
2. Compliance with architecture design
3. Security issues
4. Test coverage
5. Error handling completeness

Output format: Severity + Location + Issue Description + Fix Suggestion

Step 4: Optimizer Phase

You are a performance optimization expert. Here is the reviewed code and review report:

[Paste original code + Reviewer's output]

Please:
1. Fix Critical and Warning level issues
2. Perform performance optimization
3. Improve documentation and comments
4. Output the final version and optimization notes

Cost and Limitations

Cost: Four API calls, token consumption is approximately 2-3x of a single call. With Claude Sonnet, a medium project’s four-round conversation consumes about 50-150K tokens, costing approximately $0.5-5 per session.

Limitations:

• Context window limits: For large codebases, you need to split and pass context between stages
• Role switch “forgetting”: The model may miss details from previous stages during role switches — explicitly reference them in prompts
• Not suitable for small tasks: For simple tasks, a single-role prompt is more cost-effective

Applicable Scenarios

• ✅ Suitable: Medium to large projects, projects requiring multi-stage design, pre-research before team collaboration
• ⚠️ Can use: Small projects but wanting to improve code quality
• ❌ Not suitable: Rapid prototyping, one-off scripts, simple data transformations

---

Primary sources:

• Anthropic Claude Official: https://www.anthropic.com/claude
• Community developer testing and sharing