tmdgusya/harness-prac/multi-agent-scan

Multi-Agent Parallel Scan

Orchestrates multiple agents in parallel to perform comprehensive codebase analysis across different dimensions simultaneously.

When to Use

Invoke this skill when the user requests:

• "Scan my codebase"
• "Analyze multiple patterns"
• "Multi-agent workflow"
• "Parallel analysis"
• "Comprehensive code review"

Workflow Pattern

Phase 1: Parallel Exploration (Independent Tasks)

Fire multiple agents simultaneously to explore different aspects. All agents start in parallel when launched in a single message:

# Scan for different code patterns in parallel
Task(subagent_type="tutorial-workflow:file-scanner", model="haiku", prompt="Find all TypeScript files with TODO comments")
Task(subagent_type="tutorial-workflow:file-scanner", model="haiku", prompt="Find all files using deprecated APIs")
Task(subagent_type="tutorial-workflow:file-scanner", model="haiku", prompt="Find all large files (>500 lines)")

Phase 2: Deep Analysis (Sequential)

After results arrive, perform deeper analysis:

Task(
  subagent_type="tutorial-workflow:code-analyzer",
  model="opus",
  prompt=f"""Analyze these findings and provide recommendations:
 
  TODOs: {todo_results}
  Deprecated APIs: {deprecated_results}
  Large files: {large_files}
 
  Provide refactoring recommendations prioritized by impact."""
)

Phase 3: External Research (Parallel)

Run parallel research while analysis happens:

Task(subagent_type="tutorial-workflow:web-researcher", model="sonnet", prompt="Search for TypeScript best practices for code organization")
Task(subagent_type="tutorial-workflow:web-researcher", model="sonnet", prompt="Find latest recommendations for handling deprecated APIs")

Best Practices

1. Always use single message for parallel calls - Multiple Task calls in one message execute concurrently
2. Match model to task complexity - Use Haiku for simple searches, Sonnet for research, Opus for analysis
3. Background long-running tasks - Use run_in_background: true for builds, installs, tests
4. Limit concurrency - Maximum 5 concurrent background tasks to prevent resource exhaustion

Concurrency Limits

• Max 5 concurrent background tasks - Claude Code's default limit
• No limit on parallel foreground tasks - As many as your message can fit
• Batch dependent work - Run independent batches in parallel, wait, then next batch

Example Session

User: "Scan my codebase using the multi-agent workflow"
 
Claude:
1. Spawns 3 file-scanner agents in parallel (TODOs, deprecated APIs, large files)
2. Collects results from all agents
3. Spawns code-analyzer for deep analysis
4. Spawns web-researcher agents for best practices research
5. Provides comprehensive report with prioritized recommendations