Daniel's Agents

Dynamic agent registry and the ask_agent dispatcher for specialized AI subagents.

Overview

Daniel's Agents is a system for registering, managing, and invoking specialized AI agents. Unlike Claude Code's built-in Task tool (which has fixed, predefined subagent types), this system allows you to:

• Register custom agents with their own MCP tools and personalities
• Route requests dynamically to any registered agent via ask_agent
• Configure agents as data - system prompts, tools, and models stored in PostgreSQL

The key insight: agents are configuration, not code. Each agent is a row in a database with its MCP servers, system prompt, and model preference. The ask_agent tool looks up the config and invokes the Claude Agent SDK.

The Agent Registry

Agents are stored in the argus.agents table:

Field	Description
id	Unique identifier (e.g., venus, sentinel)
name	Display name
description	What the agent does
category	Agent type: design, qa, research, etc.
command	CLI command for the agent
model	Claude model to use
config	JSONB with mcp_servers, system_prompt, permission_mode
is_active	Whether the agent is available

Built-in Agents

Two agents are seeded by default:

Venus (Design Specialist)

{
  "mcp_servers": {
    "magic": {"command": "npx", "args": ["-y", "@anthropic/magic-mcp"]},
    "context7": {"command": "npx", "args": ["-y", "@upstash/context7-mcp@latest"]}
  },
  "system_prompt": "You are Venus, a design specialist. Help users create beautiful, accessible, and responsive UI components.",
  "permission_mode": "bypassPermissions"
}

Venus has access to:

• magic-mcp: 21st.dev component library for finding UI components
• context7: Documentation lookup for frameworks

Sentinel (QA Specialist)

{
  "mcp_servers": {
    "playwright": {"command": "npx", "args": ["-y", "@anthropic/mcp-playwright"]},
    "context7": {"command": "npx", "args": ["-y", "@upstash/context7-mcp@latest"]}
  },
  "system_prompt": "You are Sentinel, a QA specialist. Be skeptical. Verify everything works. Use TodoWrite to create checklists.",
  "permission_mode": "bypassPermissions"
}

Sentinel has access to:

• playwright-mcp: Browser automation for E2E testing
• context7: Documentation lookup

The ask_agent Tool

The core dispatcher tool that routes requests to any registered agent:

ask_agent(agent_id="venus", request="Design a login form with email and password")

How It Works

┌─────────────────────────────────────────────────────────────┐
│                    Claude Code (Main Agent)                   │
│                                                               │
│  "I need a React component for user settings"                │
│                                                               │
└──────────────────────────┬────────────────────────────────────┘
                           │ ask_agent(agent_id="venus", request="...")
                           ▼
┌─────────────────────────────────────────────────────────────┐
│                    Janus MCP Server                          │
│                                                               │
│  1. Fetch agent config from argus.agents                     │
│  2. Build ClaudeAgentOptions from stored config              │
│  3. Invoke Claude Agent SDK with config                      │
│  4. Stream response back                                      │
│                                                               │
└──────────────────────────┬────────────────────────────────────┘
                           │ Agent SDK invocation
                           ▼
┌─────────────────────────────────────────────────────────────┐
│                    Venus Subagent                            │
│                                                               │
│  Brain: Claude Sonnet                                        │
│  Tools: magic-mcp, context7                                  │
│  Prompt: "You are Venus, a design specialist..."             │
│                                                               │
│  Returns: Component code, usage examples, dependencies       │
│                                                               │
└─────────────────────────────────────────────────────────────┘

Parameters

Parameter	Type	Required	Description
agent_id	string	Yes	Agent ID to invoke (e.g., venus, sentinel)
request	string	Yes	The task or question for the agent
context	string	No	Additional context to include

Example Response

{
  "agent_id": "venus",
  "agent_name": "Venus",
  "response": "Here's a login form component using shadcn/ui..."
}

Agent Management Tools

List Agents

janus_agent_list(category="design", include_inactive=false)

Returns all registered agents, optionally filtered by category.

Get Agent Details

janus_agent_get(agent_id="venus")

Returns full agent configuration including MCP servers and system prompt.

Register New Agent

janus_agent_register(
  agent_id="athena",
  name="Athena",
  command="athena",
  description="Research specialist - documentation and best practices",
  category="research",
  model="claude-sonnet-4-20250514",
  config={
    "mcp_servers": {
      "context7": {"command": "npx", "args": ["-y", "@upstash/context7-mcp@latest"]}
    },
    "system_prompt": "You are Athena, a research specialist. Find documentation, cite sources, distinguish facts from opinions.",
    "permission_mode": "bypassPermissions"
  }
)

Update Agent

janus_agent_update(
  agent_id="venus",
  model="claude-opus-4-20250514",
  is_active=true
)

Delete Agent

janus_agent_delete(agent_id="athena")

Reeves Personas

The agent registry is also the foundation for Reeves, Daniel's personal life management system. Reeves uses personas - specialized agents for different domains:

Persona	Role	Domain
Reeves	Chief of Staff	Task management, scheduling
Penn	Historian	Journals, reflections, mood tracking
Finch	Retriever	Read-only search across all domains
Hartley	Trainer	Fitness metrics, health logs
Sterling	CFO	Financial advice (reads Tally data)
Tally	Bookkeeper	Double-entry accounting ledger

These personas share the same PostgreSQL backend as infrastructure agents but operate on personal data:

• service_id = djs-personal for personal tasks
• journals table for Penn
• health_logs table for Hartley
• tally.* schema for financial ledger

The architecture principle: personas are configuration, not separate codebases. One unified system, specialized behaviors.

API Endpoints

The agent registry is also accessible via REST API:

Method	Endpoint	Description
GET	/api/v1/agents	List all agents
GET	/api/v1/agents/{id}	Get agent details
POST	/api/v1/agents	Register new agent
PATCH	/api/v1/agents/{id}	Update agent
PATCH	/api/v1/agents/{id}/config	Update agent config section
DELETE	/api/v1/agents/{id}	Delete agent
POST	/api/v1/agents/init	Initialize table with defaults

Comparison: Task Tool vs ask_agent

Capability	Claude Code Task Tool	ask_agent
Spawn subagents	Yes	Yes
Custom MCP servers	No (fixed types)	Yes (any)
Custom system prompt	No	Yes
Model selection	Limited	Full
Tool restrictions	No	Via config
Agent as data	No (hardcoded)	Yes (database)
Dynamic registration	No	Yes

When to Use Each

Use Task Tool when:

• You need basic exploration (Explore)
• You want implementation planning (Plan)
• The predefined types fit your needs

Use ask_agent when:

• You need custom MCP tools (magic-mcp, playwright, etc.)
• You want specialized personalities (Venus, Sentinel, etc.)
• You're building reusable agent patterns
• You need the agent configuration to be dynamic

Creating Your Own Agent

Step 1: Define the Specialty

What does this agent do better than a general-purpose agent?

Agent	Specialty	Key Tools
Venus	UI/UX design	magic-mcp, context7
Sentinel	QA testing	playwright, context7
Athena	Research	context7, web search
Hermes	API integration	HTTP tools

Step 2: Choose MCP Tools

Select tools that match the specialty. Keep it focused - fewer tools = more focused agent.

Step 3: Write the System Prompt

Define personality and constraints:

You are [Name], a [specialty] specialist.

Your role:
1. [Primary responsibility]
2. [Secondary responsibility]

You do NOT:
- [Constraint 1]
- [Constraint 2]

Return [expected output format].

Step 4: Register the Agent

janus_agent_register(
  agent_id="my-agent",
  name="My Agent",
  command="my-agent",
  description="What this agent does",
  category="category",
  config={
    "mcp_servers": {...},
    "system_prompt": "...",
    "permission_mode": "bypassPermissions"
  }
)

Step 5: Invoke It

ask_agent(agent_id="my-agent", request="Do the thing")

See Also

• Custom Agents - The ClaudeBox subagent pattern
• Agent HTTP API - Running agents programmatically
• Janus - Infrastructure MCP tools