Multi-Agent Generator

A powerful low-code/no-code tool that transforms plain English instructions into fully configured multi-agent AI teams — no scripting, no complexity. Powered by LiteLLM for provider-agnostic support (OpenAI, WatsonX, Ollama, Anthropic, etc.) with both a CLI and an optional Streamlit UI.

What's New in v1.0.0

• Tool Auto-Discovery & Generation - 15+ pre-built tools + natural language tool creation
• Multi-Agent Orchestration Patterns - Supervisor, Debate, Voting, Pipeline, MapReduce
• Evaluation & Testing Framework - Auto-generated tests + output quality metrics
• CLI Support - Full CLI commands for tools, evaluation, and orchestration

---

Features

Agent Generation

• Generate agent code for multiple frameworks:

  • CrewAI: Structured workflows for multi-agent collaboration
  • CrewAI Flow: Event-driven workflows with state management
  • LangGraph: LangChain's framework for stateful, multi-actor applications
  • Agno: Agno framework for Agents Team orchestration
  • ReAct (classic): Reasoning + Acting agents using AgentExecutor
  • ReAct (LCEL): Future-proof ReAct built with LangChain Expression Language (LCEL)

• Provider-Agnostic Inference via LiteLLM:

  • Supports OpenAI, IBM WatsonX, Ollama, Anthropic, and more
  • Swap providers with a single CLI flag or environment variable

• Flexible Output:

  • Generate Python code
  • Generate JSON configs
  • Or both combined

Tool Auto-Discovery & Generation (NEW!)

Create tools for your agents using plain English — no coding required:

from multi_agent_generator.tools import ToolRegistry, ToolGenerator

# Browse 15+ pre-built tools across 10 categories
registry = ToolRegistry()
web_tools = registry.list_by_category("web_search")
all_tools = registry.list_all()

# Generate custom tools from natural language
generator = ToolGenerator()
tool = generator.generate_from_description("Create a tool that fetches weather data for a city")
print(tool.code)  # Ready-to-use Python code!

Pre-built Tool Categories:

Category	Examples
Web Search	Google search, web scraper
File Operations	Read, write, list files
Data Processing	CSV parser, JSON transformer
Code Execution	Python executor, shell runner
API Integration	REST client, webhook handler
Database	SQL query, document store
Communication	Email sender, Slack notifier
Math	Calculator, statistics
Text Processing	Summarizer, translator
Image Processing	Resizer, format converter

Multi-Agent Orchestration Patterns (NEW!)

Choose from 5 battle-tested patterns to coordinate your agents:

from multi_agent_generator.orchestration import Orchestrator, PatternType

orchestrator = Orchestrator()

# Generate orchestrated system from description
result = orchestrator.generate_from_description(
    "I need a research team where a manager delegates to specialists"
)
print(result["code"])  # Complete LangGraph/CrewAI code!

# Or configure manually
config = orchestrator.create_pattern_config(
    pattern_type=PatternType.SUPERVISOR,
    agents=["researcher", "writer", "reviewer"],
    task_description="Analyze market trends"
)

Available Patterns:

Pattern	Use Case	How It Works
Supervisor	Delegating tasks to specialists	Central coordinator routes work
Debate	Reaching consensus	Agents discuss & refine answers
Voting	Democratic decisions	Agents vote on best response
Pipeline	Sequential processing	Chain of specialized steps
MapReduce	Parallel processing	Split, process, aggregate

Evaluation & Testing Framework (NEW!)

Auto-generate tests and evaluate agent quality:

from multi_agent_generator.evaluation import TestGenerator, AgentEvaluator

# Generate pytest test suites automatically
test_gen = TestGenerator()
test_suite = test_gen.generate_test_suite(
    agent_config=your_config,
    test_types=["unit", "integration", "e2e"]
)
test_suite.save("tests/")  # Ready to run with pytest!

# Evaluate agent output quality
evaluator = AgentEvaluator()
result = evaluator.evaluate(
    agent_output="The analysis shows...",
    expected_output="Market trends indicate...",
    task_description="Analyze Q4 sales data"
)
print(result.overall_score)  # 0.0 - 1.0
print(result.metrics)  # relevance, completeness, coherence, accuracy

Test Types:

• Unit Tests - Individual component testing
• Integration Tests - Multi-agent interaction
• End-to-End Tests - Full workflow validation
• Performance Tests - Response time & throughput
• Reliability Tests - Error handling & recovery
• Quality Tests - Output quality metrics

Streamlit UI

• Interactive prompt entry
• Framework selection
• Tool discovery & generation (NEW!)
• Orchestration pattern configuration (NEW!)
• Evaluation & testing dashboard (NEW!)
• Config visualization
• Copy or download generated code

---

Installation

Basic Installation

pip install multi-agent-generator

---

Prerequisites

• At least one supported LLM provider (OpenAI, WatsonX, Ollama, etc.)

• Environment variables setup:

  • OPENAI_API_KEY (for OpenAI)
  • WATSONX_API_KEY, WATSONX_PROJECT_ID, WATSONX_URL (for WatsonX)
  • OLLAMA_URL (for Ollama)
  • Or a generic API_KEY / API_BASE if supported by LiteLLM

• Be aware Agno only works with OPENAI_API_KEY without tools for Now, and will be expanded for further API's and tools in the future.

You can freely switch providers using --provider in CLI or by setting environment variables.

---

Usage

Command Line

Basic usage with OpenAI (default):

multi-agent-generator "I need a research assistant that summarizes papers and answers questions" --framework crewai

Using WatsonX instead:

multi-agent-generator "I need a research assistant that summarizes papers and answers questions" --framework crewai --provider watsonx

Using Agno:

multi_agent_generator "build a researcher and writer" --framework agno --provider openai --output agno.py --format code

Using Ollama locally:

multi-agent-generator "Build me a ReAct assistant for customer support" --framework react-lcel --provider ollama

Save output to a file:

multi-agent-generator "I need a team to create viral social media content" --framework langgraph --output social_team.py

Get JSON configuration only:

multi-agent-generator "I need a team to analyze customer data" --framework react --format json

Tool Generation via CLI

Generate custom tools from natural language:

# Generate a custom tool
multi-agent-generator --tool "Create a tool to fetch weather data from an API"

# Save to file
multi-agent-generator --tool "Create a web scraper tool" --output scraper_tool.py

# List all available tools
multi-agent-generator --list-tools

# List tools by category
multi-agent-generator --list-tools --tool-category api_integration

Evaluation via CLI

Evaluate agent outputs directly from the command line:

# Basic evaluation
multi-agent-generator --evaluate --query "What is AI?" --response "AI is artificial intelligence..."

# With expected output for accuracy scoring
multi-agent-generator --evaluate \
  --query "Summarize machine learning" \
  --response "ML is a subset of AI that learns from data" \
  --expected "Machine learning is an AI technique" \
  --threshold 0.8

# Save results to file
multi-agent-generator --evaluate --query "Test" --response "Response" --output results.json

Orchestration via CLI

Create orchestrated multi-agent systems:

# Get pattern suggestion from description
multi-agent-generator --orchestrate "I need agents to debate and reach consensus"

# Generate code for a specific pattern
multi-agent-generator --pattern supervisor --framework langgraph --output supervisor.py

# List all available patterns
multi-agent-generator --list-patterns

# Customize number of agents
multi-agent-generator --pattern voting --num-agents 5 --framework crewai

Streamlit UI

Launch the interactive web interface:

streamlit run streamlit_app.py

Navigate between pages:

• Agent Generator - Generate agent code from natural language
• Tool Discovery - Browse and create tools
• Orchestration Patterns - Configure multi-agent coordination
• Evaluation & Testing - Generate tests and evaluate outputs

---

Examples

Research Assistant

I need a research assistant that summarizes papers and answers questions

Content Creation Team

I need a team to create viral social media content and manage our brand presence

Customer Support (LangGraph)

Build me a LangGraph workflow for customer support

Orchestrated Team (NEW!)

from multi_agent_generator.orchestration import Orchestrator

orchestrator = Orchestrator()
result = orchestrator.generate_from_description(
    "Build a content team with a supervisor managing writers and editors"
)

---

Frameworks

CrewAI

Role-playing autonomous AI agents with goals, roles, and backstories.

CrewAI Flow

Event-driven workflows with sequential, parallel, or conditional execution.

LangGraph

Directed graph of agents/tools with stateful execution.

Agno

Role-playing Team orchestration AI agents with goals, roles, backstories and instructions.

ReAct (classic)

Reasoning + Acting agents built with AgentExecutor.

ReAct (LCEL)

Modern ReAct implementation using LangChain Expression Language — better for debugging and future-proof orchestration.

---

LLM Providers

OpenAI

State-of-the-art GPT models (default: gpt-4o-mini).

IBM WatsonX

Enterprise-grade access to Llama and other foundation models (default: llama-3-70b-instruct).

Ollama

Run Llama and other models locally.

Anthropic

Use Claude models for agent generation.

...and more, via LiteLLM.

---

API Reference

Tools Module

from multi_agent_generator.tools import (
    ToolRegistry,      # Browse pre-built tools
    ToolGenerator,     # Generate custom tools
    ToolCategory,      # Tool category enum
    ToolDefinition,    # Tool data class
)

Orchestration Module

from multi_agent_generator.orchestration import (
    Orchestrator,      # High-level orchestration interface
    PatternType,       # Pattern type enum
    SupervisorPattern, # Supervisor pattern
    DebatePattern,     # Debate pattern
    VotingPattern,     # Voting pattern
    PipelinePattern,   # Pipeline pattern
    MapReducePattern,  # MapReduce pattern
)

Evaluation Module

from multi_agent_generator.evaluation import (
    TestGenerator,     # Auto-generate test suites
    TestCase,          # Individual test case
    TestSuite,         # Collection of tests
    AgentEvaluator,    # Evaluate agent outputs
    EvaluationResult,  # Evaluation results
    Benchmark,         # Performance benchmarking
)

---

License

MIT

Maintainers: Nabarko Roy

Made with love. If you like star the repo and share it with AI Enthusiasts.