I'm excited to announce - I'm writing a book titled "Multi-Agent Systems with AutoGen" published by Manning Publications Co. As of today, the first chapters are available through Manning's Early Access Program (MEAP)!

Recently, I've received numerous inquiries from software engineers, product developers, and engineering leaders about multi-agent systems. Common questions include how to build/configure them, when to use them, and how to integrate/deploy them in end-user applications. This book aims to address these questions and share insights gathered over the past year while contributing to the AutoGen open-source framework and developing tools like AutoGen Studio.

What to Expect in the Book

In This Book, You'll Learn About:

• Core components of multi-agent systems and their implementation using tools like AutoGen and AutoGen Studio
• User experience design guidelines for integrating multi-agent systems into end-user applications
• Designing and building generic agents that solve problems by interacting with interfaces (e.g., carrying out actions on webpages, mobile apps, desktop apps)
• Evaluating multi-agent systems using off-the-shelf benchmarks like GAIA, GPTQ, and SWEBench, as well as creating custom benchmarks
• Optimizing performance through agent parameter tuning, small model fine-tuning, parallel processing, dedicated planning agents, and more
• Deep dives into several use cases, such as multi-agent workflows for data analysis, customer service tasks, and creativity

Through Manning's Early Access Program, you'll receive the first chapters immediately and access a forum where I can directly answer your questions. Later chapters will be released progressively.

Why Write About Multi-Agent Systems?

The field of building autonomous or semi-autonomous multi-agent systems (enabled by generative AI models) is relatively new. As such, there are no well-defined standards or well-known design patterns yet. We are essentially in a phase of learning through experimentation, discovering the right abstractions, building appropriate frameworks, and converging on effective protocols. This has been my experience while building AutoGen Studio - a widely used low-code/no-code UI for building and testing multi-agent workflows (with >21k downloads per month on PyPI).

Despite the field's nascence, there has been intense interest, with industry leaders like Andrew Ng mentioning agentic systems and multi-agent collaboration patterns as the next frontier for generative AI.

In my opinion, there are several potentially compelling reasons why multi-agent systems are important:

• Solving new types of tasks (complex tasks): Some problems have defied automation or traditional software solutions so far. These problems are typically long-running, involve multiple steps that often require diverse expertise, necessitate planning/coordination across these steps, and frequently exist in dynamic environments. They often require human oversight to manage each step until successful completion. In the book, I refer to these as complex tasks. Example: Back-office staff that triage emails, scan for invoice attachments, extract data from each invoice, and enter them into four disparate internal systems.

• Helping users save time: If we can successfully achieve the first point above, we can save a significant amount of human effort and reduce tedium. This is how organizations, through their products, generate revenue.

• Creating disruptive new digital interfaces: Until now, we've had custom-tailored software for various tasks - an app for creating charts (e.g., matplotlib), a separate app for creating and editing documents (Word, Google Docs, etc.), a separate app for exercise and training etc. A truly sophisticated, general-purpose multi-agent system could accomplish all of this, adapting to provide the right interface given the user's context and the task at hand. This could change how software developers build apps (see an excerpt diagram from the book below), how users consume these apps. All potentially disruptive, and worth careful discussions and experimentation.

• Thinking through responsible AI considerations: Well, when systems can act on our behalf, the costs of interacting with these systems can change. Agents downloading files for us might be low cost. Agents canceling our meetings might be medium cost. Agents that can delete medical records - high cost.This means that we must devote careful discussion to understanding this space, finding the right language and structure to understand and minimize these costs.

To be clear, the best multi-agent systems we have today still fall short of fully realizing these three benefits. However, like every area of progress, what we have learned from current experimentation, coupled with rapid advancements in reasoning capabilities (and reducing costs), make this the right time to build and experiment!

Who is the Book For?

This book is ideal for you if you are:

• A software developer looking to expand your skills in AI-driven applications and leverage multi-agent systems for complex problem-solving.
• An AI practitioner interested in understanding and applying multi-agent architectures within existing AI projects.
• A system architect or engineer aiming to design more efficient and sophisticated AI solutions.
• A technical product manager seeking a deeper understanding of multi-agent AI technologies.
• A designer looking to create intuitive user interfaces based on a multi-agent solution stack.

About the Authors

Victor Dibia

Principal Research Software Engineer, Microsoft Research

Victor Dibia (PhD) is a Principal Research Software Engineer at Microsoft Research, where he has contributed to projects like GitHub Copilot that serves millions of customers. He is the creator of LIDA, a widely used tool for automated visualizations using generative AI models, and AutoGen Studio , a low-code tool for prototyping multi-agent applications. Victor holds a PhD in Information Systems from City University of Hong Kong and an MSc in Computer Science from Carnegie Mellon University.

Victor Dibia, PhD

FAQs

• Is the book limited to AutoGen?
  No. Across the book, the goal is to teach core principles and AutoGen is used as a tool to illustrate these principles. We expect that these principles can be implemented in any sufficiently expressive multi-agent framework

• Is there a discount code for the book?
  Yes! Use the code 'mldibia' for a 50% discount valid until August 5th.

• Can I use multiple models with the agents described in the book?
  Yes! AutoGen natively supports multiple generative AI model providers - OpenAI , Microsoft Azure OpenAI, Google gemini models, Anthropic , Cohere etc. And ofcourse you can use local hosted models from HuggingFace.

• Is there a GitHub repository for the book?
  Yes! The GitHub repository is available at GitHub Repository Link

• How do I stay updated on book progress?
  You can keep up with updates at the Multi-Agent Book website.

Acknowledgements

None of this work would be possible without all the engaged maintainers and contributors to the AutoGen OSS project. Thank you!