LLM-MAR

LLM-MAR (Multi Agent Reasoning) is a compact CLI that creates LLM agents, lets them debate, answers questions, and builds workflows.

Table of Contents

• Installation
• Quick Start
• Usage
• YAML Structure

Installation

Option 1: Install from npm (Recommended)

npm install -g llm-mar

Option 2: Install from source

1. Clone the repository:

   git clone https://github.com/llm-mar/llm-mar.git
   cd llm-mar

2. Install dependencies:

   npm install

3. Set up your OpenAI API key:

   export OPENAI_API_KEY=your_api_key_here

Quick Start

Get started with LLM-MAR in minutes! Create intelligent agents, form teams, and orchestrate debates—all through simple YAML configurations.

1. Create Your First Agent

Let's create an AI assistant specialized in creative writing:

llm-mar create agent writer --model gpt-4 --goal "To write engaging stories" --role "Creative Writer" --system-prompt "You are a talented fiction writer with a vivid imagination." --instructions "Use descriptive language,Build suspense,End with a twist" --output text

This generates a YAML configuration file that defines your agent's personality and capabilities.

2. See the YAML Structure

The command above creates default/writer.yaml with this structure:

version: '1.0'
kind: Agent
metadata:
  name: writer
  description: An agent named writer
spec:
  id: writer
  model: gpt-4
  goal: To write engaging stories
  role: Creative Writer
  system_prompt: You are a talented fiction writer with a vivid imagination.
  instructions:
    - Use descriptive language
    - Build suspense
    - End with a twist
  output: text

Key Components:

• metadata: Basic info and description
• spec: The agent's configuration including model, role, and behavior
• instructions: List of guidelines for the AI's responses

3. Run Your Agent

Now let's use the agent to generate a story:

llm-mar run default/writer.yaml --input "Write a short story about a mysterious old clock in an antique shop."

The agent will respond with a creative story based on its configuration!

4. Build a Team

Combine multiple agents for collaborative tasks:

llm-mar create team story_team --agents "default/writer.yaml,default/editor.yaml" --output text

5. Set Up a Debate

Create engaging debates between agents:

llm-mar create debate tech_debate --agents "default/optimist.yaml,default/pessimist.yaml" --judges "default/judge.yaml" --input "Will AI replace human jobs?" --output text

Usage

After installation, use the llm-mar command (or npx llm-mar if installed locally).

Creating Agents

Create a new agent YAML file:

llm-mar create agent myagent --model gpt-4 --goal "To answer questions" --role "Assistant" --system-prompt "You are a helpful assistant." --instructions "Think step by step,Answer clearly" --output text

This creates default/myagent.yaml.

Creating Teams

Create a team of agents:

llm-mar create team myteam --agents "default/agent1.yaml,default/agent2.yaml" --output text

Running Agents

Run an agent with input:

llm-mar run default/myagent.yaml --input "What is the capital of France?"

Running Teams

Run a team with a specific agent:

llm-mar run default/myteam.yaml --agent agent1 --input "Discuss this topic"

Running Debates

Run a debate (assuming you have a debate YAML):

llm-mar run default/debate1.yaml

YAML Structure

LLM-MAR uses YAML files to define agents, teams, and debates. Here's a quick overview of the structure:

Agent YAML

version: '1.0'
kind: Agent
metadata:
  name: Scientist
  description: A scientific analysis agent
spec:
  id: scientist
  model: gpt-4
  goal: To provide scientific explanations
  role: Research Scientist
  system_prompt: You are an expert scientist...
  instructions:
    - Use evidence-based reasoning
    - Explain complex concepts simply
  output: text

Team YAML

version: '1.0'
kind: Team
metadata:
  name: Research Team
  description: Collaborative research team
spec:
  agents:
    - default/scientist.yaml
    - default/analyst.yaml
  output: text

Debate YAML

version: '1.0'
kind: Debate
metadata:
  name: Ethics Debate
  description: Debate on AI ethics
spec:
  method: majority_vote
  input: Should AI have rights?
  judges:
    - default/judge.yaml
  agents:
    - default/pro.yaml
    - default/con.yaml
  output: text

For detailed configuration options and advanced features, see docs/yaml-structure-guide.md.