⚡ Lightning Agents ⚡

Dynamic agent instantiation using the Factory-of-Factories pattern with Claude Agent SDK.

Built for the Austin AI MUG Lightning Talk. 🎤

---

🤔 What is Lightning Agents?

Lightning Agents demonstrates a pattern for dynamically loading, instantiating, and even generating AI agents from declarative JSON definitions. Instead of hardcoding agent configurations, you define blueprints that get transformed into factory functions at runtime.

😤 The Problem

Building AI agents typically involves:

• 📁 Hardcoded system prompts scattered across files
• 🔗 Tightly coupled agent definitions and execution logic
• 🚫 No standardized way to add new agents without code changes
• 📋 Manual configuration duplication when agents share patterns

⚡ The Solution

Factory-of-Factories: A registry that loads agent definitions from JSON, builds factory functions for each, and provides a unified interface for instantiation with runtime context injection.

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│ Definition  │ ──▶ │   Factory   │ ──▶ │  Registry   │ ──▶ │  Instance   │
│   (JSON)    │     │  (Callable) │     │  (Unified)  │     │   (Ready)   │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘

---

🚀 Quick Start

# Clone and setup
cd lightning-agents
uv sync

# Copy env template and configure
cp .env.example .env
# Edit .env with your SEARXNG_URL

# List available agents
lightning list

# Run an agent
lightning run basic_helper "What is the factory pattern?"

# Create a new agent with the architect
lightning architect "code reviewer for Python security"

---

💻 CLI Commands

⚡ List Agents

lightning list

Shows all available agents with descriptions and model info.

⚡ Run Agent

lightning run <agent_id> "<prompt>"

Examples:

lightning run basic_helper "Explain dependency injection"
lightning run aimug_researcher "What RAG tutorials does AIMUG have?"
lightning run lab_finder "Find labs about LangGraph"

⚡ Create New Agent

lightning architect "<task description>"

The architect agent generates a new agent definition, saves it to agents.json, and makes it immediately available.

lightning architect "meeting notes summarizer for action items"
# New agent created: meeting_summarizer
lightning run meeting_summarizer "Summarize: [transcript]"

---

🤖 Available Agents

Agent	Description	Model	Tools
basic_helper	General Q&A assistant	haiku	-
research_assistant	Structured research summaries	sonnet	web_search
python_doc_writer	Python function documentation	haiku	-
architect	🏗️ Designs new agent definitions (Voyager-style)	sonnet	db_agents (CRUD)
tool_architect	🔧 Designs new custom tools	sonnet	db_tools (CRUD)
aimug_researcher	Searches AIMUG content (GitHub, docs, YouTube)	sonnet	web_search, url_read
lab_finder	Finds AIMUG labs by topic	haiku	web_search
git_commit_writer	✍️ Writes conventional commit messages	haiku	-
presentation_slide_writer	🎨 Creates and manages PPTX presentations	sonnet	slides (CRUD), generate_pptx, Read, Bash
paper_researcher	📄 Researches papers, downloads PDFs	sonnet	web_search, url_read, download_pdf

---

🏭 Factory vs Factory-of-Factories

Simple Factory

def create_agent(config):
    return Agent(config)

Factory-of-Factories (This Project) ⚡

# Definition → Factory → Registry → Instance
registry = AgentRegistry.from_json("agents.json")  # Builds ALL factories
agent = registry.create("researcher", {"topic": "AI"})  # Gets instance

The registry is the factory-of-factories — it produces factory functions from definitions, then uses those factories to produce instances.

---

🏗️ Architect Agents

The "Architect Agent" pattern: an agent that generates new agent definitions.

# Before: 6 agents
lightning list

# Use architect to create a new one
lightning architect "code reviewer for security vulnerabilities"

# After: 7 agents - new one saved to agents.json
lightning list

# Use it immediately
lightning run security_reviewer "Review this auth code..."

This enables self-expanding agent systems where the AI itself designs specialized agents for new tasks. 🤯

---

🎮 Voyager Inspiration

This project draws inspiration from Voyager, an AI agent that plays Minecraft by building a skill library that grows over time. Instead of hardcoded behaviors, Voyager learns new skills and stores them for reuse.

Lightning Agents applies this concept to agent systems:

• Skill Library → Agent Registry: Agents stored as reusable definitions
• Learning New Skills → Architect Agent: Creates new agents on demand
• Tool Acquisition → Tool Architect: Creates new tools when needed

The result: a system that grows organically based on actual needs, not pre-planned capabilities.

---

🧪 Hypotheses

H1: Declarative > Imperative for Agent Configuration

JSON definitions separate what an agent is from how it runs. Easier to version, diff, and review.

H2: Runtime Context Injection Enables Reusability

The same definition can serve multiple contexts by injecting different runtime parameters.

H3: Architect Agents Enable Organic Growth

Instead of manually writing every agent definition, let Claude design agents for new tasks. The system grows based on actual needs.

---

📁 Project Structure

lightning-agents/
├── .env.example              # Environment template
├── .env                      # Your config (gitignored)
├── pyproject.toml
├── README.md
├── CLAUDE.md                 # Developer notes
├── db/                       # Data (decoupled from source)
│   ├── agents.json           # Agent blueprints
│   └── tools.json            # Custom tool definitions
├── src/lightning_agents/     # Main package
│   ├── __init__.py
│   ├── cli.py                # CLI entry point
│   ├── runner.py             # Agent execution with MCP
│   ├── registry.py           # Factory-of-Factories pattern
│   ├── agent_factory.py      # Definition → Instance
│   ├── mcp_config.py         # MCP server configs
│   └── tools/                # Custom MCP tools
│       ├── download_pdf.py   # PDF download tool
│       ├── db_agents.py      # Agent CRUD operations
│       ├── db_tools.py       # Tool CRUD operations
│       └── presentation.py   # Slide manipulation tools
└── presentation/             # PPTX slide generator
    ├── generate_slides.py
    ├── slide_content.py
    ├── styles.py
    └── output/
        └── lightning-agents.pptx

---

🎨 Generating Slides

Use the presentation_slide_writer agent to manage slides:

# List current slides
lightning run presentation_slide_writer "List the slides"

# Add a new slide
lightning run presentation_slide_writer "Add a bullets slide about MCP integration"

# Generate PPTX and PDF
lightning run presentation_slide_writer "Generate the presentation"

# Output: presentation/output/lightning-agents.pptx + .pdf

Edit presentation/slide_content.py directly for bulk changes. Supports **bold** and `code` markup.

---

⚙️ Configuration

Environment Variables

Copy .env.example to .env and configure:

# SearXNG MCP server URL (for web search agents)
SEARXNG_URL=http://localhost:8888

MCP Tools

Agents can use MCP tools by declaring them in their tools array:

SearXNG (Web Search):

• mcp__searxng__searxng_web_search - Web search
• mcp__searxng__web_url_read - Read web page content

Custom Tools (Built-in MCP Server):

• mcp__custom-tools__download_pdf - Download PDFs from URLs
• mcp__custom-tools__db_list_agents / db_get_agent / db_create_agent / db_update_agent / db_delete_agent - Agent CRUD
• mcp__custom-tools__db_list_tools / db_get_tool / db_create_tool / db_update_tool / db_delete_tool - Tool CRUD
• mcp__custom-tools__list_slides / add_slide / update_slide / delete_slide / generate_pptx - Presentation management

SDK Primitives: Agents can also use built-in Claude SDK tools: Read, Write, Edit, Bash, Grep, Glob, WebFetch, WebSearch

---

📋 Requirements

• Python 3.13+
• uv package manager
• SearXNG instance (for search-enabled agents)

---

📄 License

MIT

---

⚡ Built for the Austin AI MUG lightning talk on dynamic agent instantiation patterns. ⚡

Agents creating agents creating agents... 🤖➡️🤖➡️🤖