15. Multi-Agent Software Engineering

15.1 Why multiple agents?

A single agent can perform many tasks, but specialization can make workflows easier to control.

Example:

Main Agent
 |
 +--> Architect
 |
 +--> Developer
 |
 +--> Tester
 |
 +--> Reviewer
 |
 +--> Security
 |
 +--> Documentation

15.2 Feature workflow

Requirement:

Add user authentication.

Step 1 — Explore

Find:
- Authentication code
- Database models
- API routes
- Existing tests

Do not modify files.

Step 2 — Architecture

Design the authentication architecture.

Do not modify files.

Return:
- Components
- Data model
- API endpoints
- Security considerations
- Tests

Step 3 — Implement

Implement the approved design.

Follow AGENTS.md.

Run relevant tests.

Step 4 — Test

Analyze authentication coverage.

Add or recommend:
- Happy path
- Invalid credentials
- Missing credentials
- Authorization failures
- Edge cases

Step 5 — Review

Review the changes.

Do not modify files.

Report correctness and maintainability findings.

Step 6 — Security

Perform a security review.

Focus on:
- Password handling
- Tokens
- Authorization
- Input validation
- Secrets

Step 7 — Documentation

Update documentation for the final implementation.

15.3 Parallel vs sequential work

Some tasks can be parallelized:

              Main Agent
             /          \
            v            v
        Tester        Security

Other tasks should be sequential:

Architecture
     |
     v
Implementation
     |
     v
Testing
     |
     v
Fixes

Do not parallelize tasks that depend on the result of earlier work.

15.4 Human in the loop

Humans should remain involved in decisions involving:

• Architecture
• Security
• Data loss
• Production
• External side effects
• Large refactors

Exercise

Design a multi-agent workflow for:

Build a payment API.

Include:

• Explorer
• Architect
• Developer
• Tester
• Security
• Reviewer
• Documentation
• Human approval points