Subagent Manager

Subagent Manager

Overview

This skill helps manage and coordinate multiple Claude Code subagents for parallel task execution. It provides patterns for launching, monitoring, and coordinating subagents effectively.

When to Use

Use this skill when:

• User requests parallel execution of multiple independent tasks
• Need to explore multiple parts of a codebase simultaneously
• Running multiple test suites or builds in parallel
• Coordinating complex multi-step workflows
• Managing long-running background tasks

Prerequisites

• Claude Code CLI available (claude command)
• Understanding of Task tool with subagent_type parameter
• Knowledge of available subagent types: Bash, general-purpose, Explore, Plan

Instructions

Step 1: Identify Parallelizable Work

Analyze the user's request to identify:

• Independent tasks that can run in parallel
• Dependencies between tasks (must run sequentially)
• Tasks that require different specializations

Step 2: Select Appropriate Subagent Types

Step 3: Launch Subagents

For parallel execution, use multiple Task tool calls in a single message.

Example parallel launch pattern:

• Task 1: Use Explore agent to find all API endpoints
• Task 2: Use Explore agent to find all database models
• Task 3: Use Bash agent to run lint checks

Step 4: Coordinate Results

When subagents return:

1. Collect and summarize outputs from all agents
2. Identify any conflicts or overlapping work
3. Synthesize findings into coherent response
4. Track agent IDs for potential follow-up (resume capability)

Step 5: Background Task Management

For long-running tasks, use run_in_background: true:

• Agent runs asynchronously
• Returns output_file path immediately
• Use Read tool or Bash tail to check progress
• Continue other work while waiting

Expected Output

• Coordinated execution of multiple subagents
• Aggregated results from parallel tasks
• Clear summary of what each agent accomplished
• Agent IDs preserved for potential resumption

Error Handling

Error: Subagent timeout

• Cause: Task too complex or resource intensive
• Fix: Break into smaller subtasks, increase timeout, or run in background

Error: Conflicting results from parallel agents

• Cause: Agents modified same files or reached different conclusions
• Fix: Run sequentially with dependencies, or reconcile manually

Error: Agent requires information from another agent

• Cause: Task has hidden dependencies
• Fix: Identify dependency, run prerequisite agent first, then dependent agent

Examples

Example 1: Parallel Codebase Exploration

User request: "Understand how authentication and authorization work in this codebase"

Subagent coordination:
1. Launch Explore agent: "Find all authentication-related files and explain the auth flow"
2. Launch Explore agent: "Find all authorization/permission-related files and explain access control"
3. Synthesize: Combine findings into comprehensive auth/authz overview

Example 2: Parallel Build and Test

User request: "Run the build and tests in parallel"

Subagent coordination:
1. Launch Bash agent (background): "Run npm run build"
2. Launch Bash agent (background): "Run npm test"
3. Monitor: Check output files for completion
4. Report: Aggregate build and test results

Example 3: Multi-Repository Analysis

User request: "Compare the API design between v2 and v3"

Subagent coordination:
1. Launch Explore agent: "Analyze v2 API structure and patterns"
2. Launch Explore agent: "Analyze v3 API structure and patterns"
3. Synthesize: Create comparison table highlighting differences

Subagent Capabilities Reference

Bash Agent

• Execute shell commands
• Git operations (commit, push, branch)
• File system operations via commands
• Process management

General-Purpose Agent

• Full tool access (except Task, Edit, Write)
• Multi-step reasoning
• Code search and analysis
• Web fetching and research

Explore Agent

• Fast codebase exploration
• File pattern matching (glob)
• Code keyword search (grep)
• Architecture understanding
• Thoroughness levels: quick, medium, very thorough

Plan Agent

• Implementation strategy design
• Architectural trade-off analysis
• Step-by-step planning
• Critical file identification

Best Practices

1. Minimize Context: Give agents focused prompts with only necessary context
2. Use Appropriate Types: Match agent type to task requirements
3. Parallel When Possible: Launch independent tasks together
4. Sequential When Needed: Respect dependencies between tasks
5. Background for Long Tasks: Use run_in_background for builds, tests
6. Preserve Agent IDs: Save IDs for potential follow-up work
7. Aggregate Clearly: Synthesize multi-agent results coherently

Resources

• Claude Code Task tool documentation
• Forgekeeper v3 skills system (v3/skills/)
• Agent orchestration patterns (frontend/server/agents/)

Notes

• Each subagent starts fresh unless resumed with an agent ID
• Background agents can be checked via output_file
• Prefer haiku model for quick, straightforward tasks
• Use sonnet/opus for complex reasoning tasks

Version History

• 1.0.0 (2025-02-02): Initial release

Converted and distributed by TomeVault — claim your Tome and manage your conversions.