Claude Code Agent Management

Claude Code Agent Management

Create and manage specialized AI subagents that work independently with their own context windows and custom instructions.

Quick Start

/agents
/agent create
# Describe agent purpose: "A marketer who can help me write amazing brand copy"
# Configure tools, model (sonnet/opus/haiku), and color
# Invoke: "Use the agent brand-copywriter to generate five headline ideas for a coffee shop website"

Core Workflow

When you need specialized expertise that should work independently:

1. Create Agent

   • Run /agent create or /agents → create new
   • Describe agent role and capabilities in plain English
   • Select tools (read-only recommended for content creation)
   • Choose model power level (opus for complex tasks, sonnet for general use)
   • Assign identifying color

2. Configure Agent Context

   • Agent gets own system prompt based on your description
   • Agent inherits from parent but works in isolation
   • Each agent maintains separate conversation history

3. Invoke Agent

   • Use natural language: "Use the agent [name] to [specific task]"
   • Agent appears in different color to show it's working
   • Multiple agents can work simultaneously on different tasks

Techniques

Agent Specialization Patterns:

• Brand copywriter → marketing content generation
• Security reviewer → code vulnerability analysis
• UX/UI specialist → interface design recommendations
• Planning agent → project structure and roadmaps

Multi-Agent Coordination:

• Spin up multiple instances of same agent type
• Run different specialized agents in parallel
• Each maintains independent context and memory

Agent vs Skills Decision:

• ALWAYS use agents when: Need isolated context, specialized expertise, independent thinking
• NEVER use agents when: Need to stack multiple capabilities in same context

Anti-Patterns

NEVER create agents for:

• Simple single-step tasks (use direct prompts)
• Tasks that need to reference current conversation context
• Quick utility functions (use skills instead)

NEVER configure agents with:

• Overly broad permissions without clear need
• Vague role descriptions (be specific about capabilities)

Edge Cases & Error Handling

Agent Access Issues:

• If agent list doesn't appear: restart Claude Code session
• If agent creation fails: simplify role description, check tool permissions

Context Management:

• Agent responses don't automatically merge with main conversation
• Copy important agent outputs to main thread for persistence
• Agents lose context when session ends (save important configurations)

Performance Considerations:

• Multiple opus agents consume significant tokens
• Use haiku/sonnet for routine tasks, reserve opus for complex reasoning

Bundled Resources Plan

• agents/templates/ - Common agent role templates and configurations
• agents/examples/ - Sample agent conversations and use cases
• scripts/agent-backup.sh - Export/import agent configurations

Source: jayalexandermg/SkillJacked — distributed by TomeVault.