mahsm 0.1.1

Multi-Agent Hyper-Scaling Methods

mahsm: Multi-Agent Hyper-Scaling Methods

A declarative framework for building, tracing, and evaluating multi-agent LLM systems

---

What is mahsm?

mahsm unifies four best-in-class libraries into a single, zero-boilerplate framework:

Component	Purpose	What mahsm Adds
DSPy	Programming LLMs with optimizable modules	Automatic LangGraph integration via @ma.dspy_node
LangGraph	Stateful, cyclical multi-agent workflows	Zero-boilerplate DSPy module wrapping
Langfuse	Deep LLM observability & tracing	Automatic instrumentation with ma.init()
EvalProtocol	Pytest-based LLM evaluation	Pre-configured harness for LangGraph apps

The mahsm Advantage

Before mahsm (Traditional approach):

# Manual state mapping, verbose boilerplate
def my_node(state: dict) -> dict:
    module = MyDSPyModule()
    result = module(input=state["input"])
    return {"output": result.output}  # Manual extraction

# Separate tracing setup
from langfuse import Langfuse
langfuse = Langfuse()
# ... complex instrumentation code ...

With mahsm (Declarative approach):

import mahsm as ma

ma.init()  # One-line tracing setup

@ma.dspy_node  # Automatic state mapping
class MyModule(ma.Module):
    def forward(self, input):
        return self.predictor(input=input)

---

⚡ Quick Start

Installation

pip install mahsm

Your First Agent (60 seconds)

import mahsm as ma
from typing import TypedDict
import dspy

# 1. Configure
lm = dspy.OpenAI(model='gpt-4o-mini')
dspy.settings.configure(lm=lm)
ma.init()  # Enable tracing

# 2. Define agent
class State(TypedDict):
    query: str
    answer: str

@ma.dspy_node
class Researcher(ma.Module):
    def __init__(self):
        super().__init__()
        self.cot = ma.dspy.ChainOfThought("query -> answer")
    
    def forward(self, query):
        return self.cot(query=query)

# 3. Build graph
workflow = ma.graph.StateGraph(State)
workflow.add_node("researcher", Researcher())
workflow.add_edge(ma.START, "researcher")
workflow.add_edge("researcher", ma.END)

graph = workflow.compile()

# 4. Run
result = graph.invoke({"query": "What is DSPy?"})
print(result["answer"])

That's it! Your agent is now:

• ✅ Running with proper state management
• ✅ Automatically traced in Langfuse
• ✅ Ready for evaluation with EvalProtocol

---

📖 Complete Tutorial

Want to build a production-ready agent with full observability and evaluation?

👉 Read the Complete Quickstart Guide

You'll learn:

• 🎯 Building complex multi-node agents
• 📊 Setting up Langfuse for tracing
• 🧪 Running systematic evaluations
• 📈 Viewing results in Langfuse & EvalProtocol UIs
• ⚡ Optimizing agents with DSPy compilers

---

🏗️ Core Features

1. @ma.dspy_node Decorator

The heart of mahsm: converts DSPy modules to LangGraph nodes automatically.

Supports two patterns:

Class Decorator

@ma.dspy_node
class MyModule(ma.Module):
    def forward(self, input1, input2):
        # Your logic here
        return self.predictor(input1=input1, input2=input2)

# Use in graph
workflow.add_node("my_node", MyModule())

Instance Wrapper

# Wrap any DSPy module instance
cot = ma.dspy.ChainOfThought("question -> answer")
node = ma.dspy_node(cot)

# Use directly
workflow.add_node("cot", node)

What it does:

• ✅ Introspects forward() parameters (excludes self)
• ✅ Extracts matching fields from LangGraph state
• ✅ Returns result as state updates (non-private fields)
• ✅ No manual state mapping needed

2. ma.init() - One-Line Tracing

handler = ma.init()

Automatically:

• ✅ Initializes Langfuse client from environment variables
• ✅ Instruments DSPy for automatic trace capture
• ✅ Returns LangChain CallbackHandler for LangGraph tracing
• ⚠️ Gracefully warns if credentials missing

3. ma.testing.PytestHarness

Bridge from LangGraph apps to EvalProtocol evaluations:

from my_agent import graph
import mahsm as ma

harness = ma.testing.PytestHarness(graph=graph)

@ma.testing.evaluation_test(
    data_loaders=harness.data_loaders,
    rollout_processor=harness.rollout_processor,
    completion_params=[{"model": "openai/gpt-4o-mini"}],
)
async def test_quality(row):
    return await ma.testing.aha_judge(row, judge_model="openai/gpt-4o", rubric="...")

---

🛠️ Development

Setup

git clone https://github.com/chimera-research/mahsm.git
cd mahsm
pip install -e .

Run Tests

python tests/test_core.py              # Unit tests
python tests/test_graph_integration.py  # Integration tests

CI/CD

• GitHub Actions: Runs tests on push/PR (Python 3.10-3.12, Linux/Mac/Windows)
• PyPI Publishing: Automatic via GitHub Releases or manual workflow dispatch

---

📊 Architecture

┌─────────────────────────────────────────────────────────────┐
│                        mahsm Framework                      │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  ┌──────────────┐      ┌──────────────┐                   │
│  │   DSPy       │─────▶│  @dspy_node  │                   │
│  │   Modules    │      │  Decorator   │                   │
│  └──────────────┘      └──────┬───────┘                   │
│                                │                            │
│                                ▼                            │
│  ┌──────────────────────────────────────────────────┐     │
│  │          LangGraph StateGraph                    │     │
│  │  ┌─────┐   ┌─────┐   ┌─────┐   ┌─────┐         │     │
│  │  │Node1│──▶│Node2│──▶│Node3│──▶│End  │         │     │
│  │  └─────┘   └─────┘   └─────┘   └─────┘         │     │
│  └──────────────────────┬───────────────────────────┘     │
│                         │                                  │
│         ┌───────────────┴───────────────┐                 │
│         │                                │                 │
│         ▼                                ▼                 │
│  ┌──────────────┐              ┌──────────────┐           │
│  │   Langfuse   │              │ EvalProtocol │           │
│  │   Tracing    │              │  Evaluation  │           │
│  └──────────────┘              └──────────────┘           │
│                                                             │
└─────────────────────────────────────────────────────────────┘

---

🤝 Contributing

We welcome contributions! Please:

1. Check existing issues or create one
2. Fork the repo and create a feature branch
3. Write tests for new functionality
4. Submit a PR with clear description

---

📄 License

MIT License - see LICENSE for details.

---

🌟 Acknowledgments

mahsm stands on the shoulders of giants:

• DSPy by Stanford NLP
• LangGraph by LangChain
• Langfuse by Langfuse Team
• EvalProtocol by Fireworks AI

---

📬 Contact

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Twitter: @chimera_research (if applicable)

---

Built with ❤️ by Chimera Research

⭐ Star us on GitHub