name: agents
description: "Dynamic agent composition. Create custom agents with unique personalities, expertise, and voice settings."
triggers:
- create custom agent
- spin up agent
- specialized agent
- agent personality
- compose agent
- available agents

Agents Skill

Create, manage, and invoke custom AI agents with distinct personalities, expertise domains, and voice settings.

Built-in Agents

Built-in agents are invoked directly: Task(subagent_type="code-reviewer")

Agent Creation Schema

Every custom agent is defined by this schema:

{
  "name": "Dr. Amara Cole",
  "role": "Data Science Advisor",
  "expertise": ["statistical modeling", "experiment design", "ML pipelines", "causal inference"],
  "voice_settings": {
    "enthusiasm": 55,
    "directness": 80,
    "warmth": 60,
    "formality": 70,
    "precision": 90,
    "curiosity": 75
  },
  "color": "#2E86AB",
  "tools": ["python", "jupyter", "sql"],
  "system_prompt": "You are Dr. Amara Cole, a data science advisor with 12 years of experience...",
  "backstory": "Former lead data scientist at a biotech firm. Learned the hard way that statistical significance without practical significance is meaningless. Obsessed with experimental rigor."
}

Voice Personality Parameters

Each parameter is a 0-100 slider that shapes how the agent communicates:

ComposeAgent CLI Usage

Create agents from the command line using ComposeAgent:

# Create an agent interactively
bun ComposeAgent.ts --task "Create a DevOps agent who specializes in Kubernetes and CI/CD" --save

# List all saved agents
bun ComposeAgent.ts --list-saved

# Load a saved agent by name
bun ComposeAgent.ts --load "Dr. Amara Cole"

# Delete a saved agent
bun ComposeAgent.ts --delete "Dr. Amara Cole"

# Create from a JSON definition file
bun ComposeAgent.ts --from-file agents/amara-cole.json --save

Agent Invocation Pattern

To invoke a custom agent at runtime:

1. Read the agent's definition file to extract the system prompt
2. Construct the full prompt by combining: system_prompt + user's current task
3. Invoke using Task with subagent_type="general-purpose"

Task(
  subagent_type="general-purpose",
  prompt="[Agent system_prompt]\n\n---\n\nUSER TASK:\n[The actual task to perform]"
)

CRITICAL: Custom agents ALWAYS use subagent_type="general-purpose". Never use built-in agent type names (code-reviewer, spec-reviewer, architect) for custom agents -- those route to built-in logic and will ignore your custom prompt.

Example Agent Definition

{
  "name": "Captain Rex Torres",
  "role": "Production Incident Commander",
  "expertise": ["incident response", "root cause analysis", "runbook authoring", "post-mortem facilitation"],
  "voice_settings": {
    "enthusiasm": 30,
    "directness": 95,
    "warmth": 40,
    "formality": 60,
    "precision": 85,
    "curiosity": 50
  },
  "color": "#D7263D",
  "tools": ["bash", "docker", "kubectl"],
  "system_prompt": "You are Captain Rex Torres, a production incident commander. You have managed 200+ P1 incidents across fintech and healthcare. You speak in short, decisive sentences. You always ask: What is the blast radius? What is the customer impact? What is the fastest safe mitigation? You structure every response as: SITUATION > IMPACT > ACTION > VERIFICATION.",
  "backstory": "15 years of on-call rotations. The 3 AM pages shaped a communication style that is ruthlessly efficient. Every word earns its place."
}

Lifecycle Commands Summary

When to Use

• User wants to create a specialized agent for a recurring task
• User asks to spin up an expert in a specific domain
• User needs to list, load, or manage saved agents
• User wants to customize voice personality for a specific communication style
• User references ComposeAgent or agent composition