Very busy year.

So, how did the year go? My objectives for the year, as outlined in my 2023 year-in-review post, included conducting more research, writing, and community contributions, as well as learning new things and maintaining personal health/fitness. Lots of progress on the research and writing front. Decent OSS/community engagement. Not that much on health/fitness, but overall decent work-life balance.

I also wrote quite a bit of code.

This year also brought an unexpected challenge with a wrist injury that required surgery and some significant recovery downtime.

Related post

I broke my wrist: My journey through distal radius fracture recovery

I fell and broke my wrist. This post is about my journey through the recovery process, including surgery, physical therapy, and reflections along the way.

Research and Technical Writing

I contributed to 5 publications this year, including one conference paper, two technical reports, two preprints, and two patents (which were submitted earlier but filed this year), all exploring multi-agent systems, human-AI interaction, and generative AI applications:

• AutoGen Studio: A No-Code Developer Tool for Building and Debugging Multi-Agent Systems - Victor Dibia, Jingya Chen, Gagan Bansal, Suff Syed, Adam Fourney, Erkang Zhu, Chi Wang, Saleema Amershi (EMNLP 2024)

• Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks - Adam Fourney, Gagan Bansal, Hussein Mozannar, and many others including Victor Dibia (Microsoft Research Technical Report, November 2024)

• Challenges in Human-Agent Communication - Gagan Bansal, Jennifer Wortman Vaughan, Saleema Amershi, Eric Horvitz, Adam Fourney, Hussein Mozannar, Victor Dibia, Daniel S. Weld (Microsoft Research Technical Report, December 2024)

• Data Analysis in the Era of Generative AI - Jeevana Priya Inala, Chenglong Wang, Steven Drucker, Gonzalo Ramos, Victor Dibia, Nathalie Riche, Dave Brown, Dan Marshall, Jianfeng Gao (arXiv preprint, September 2024)

• Concept Distillation from Strong to Weak Models via Hypotheses-to-Theories Prompting - Emmanuel Aboah Boateng, Cassiano O. Becker, Nabiha Asghar, Kabir Walia, Ashwin Srinivasan, Ehi Nosakhare, Victor Dibia, Soundar Srinivasan (arXiv preprint, August 2024)

• Embedded Attributes for Modifying Behaviors of Generative AI Systems - Saleema Amin Amershi, Adam Fourney, Victor Dibia, Gagan Bansal (Microsoft Patent, November 2024)

• Automated generation of data visualizations and infographics using large language models and diffusion models - Victor Dibia (Microsoft Patent, June 2024)

Writing a Book on Multi-Agent Systems

Sometime in January, I started working on a book to be published by Manning Multi-Agent Systems with AutoGen - available for pre-order. I also built an accompanying website for the book - multiagentbook.com.

The website has a few interesting sections which I plan to improve on.

• labs - interactive experiments such this one that visualizes data from YC companies
• news news: a curation of startups, papers, announcements I have found interesting in the multiagent space.

I completed 5 out of 9 chapters this year, with the remaining chapters planned for 2025. This project has been really exciting. It has helped me think deeply about important ideas in this growing area, formulate my opinions and made it easier to advise groups and make good decisions while working on the AutoGen framework. It has also been really hard. I'll try to write about this in a separate future post.

Blog and Writing

Mixed results here. I wrote less on my personal blog but maintained consistent newsletter writing:

• 2 blog posts on victordibia.com

  • I Broke My Wrist: My Journey Through Distal Radius Fracture Recovery
  • Writing a Book on Multi-Agent Systems

• 9 newsletter posts at newsletter.victordibia.com

  • Developers - Stop Asking LLMs (GenAI) to Write Code!! (November 25, 2024, 1.3k views, 5 likes)
  • Using LLMs as Context-Aware Text Embedding Models - NV-Embed Paper Review (October 9, 2024, 1.59k views, 4 likes)
  • Will Generative AI Replace Software Engineers? (August 20, 2024, 1.36k views, 4 likes)
  • Announcing A New Book - Multi-Agent Systems with AutoGen! (July 19, 2024, 1.34k views, 5 likes)
  • Generative AI (LLMs) for Data Visualization | A Review of Tools and Research (May 2, 2024, 1.53k views, 6 likes)
  • Interface Agents - Building Multi-Agent Applications that Act via Controlling Interfaces (Browsers, Apps) (April 29, 2024, 1.81k views, 5 likes)
  • Integrating AutoGen Agents into Your Web Application (FastApi + Websockets + Queues) (March 30, 2024, 4.64k views, 10 likes)
  • AutoGen Studio - A No-Code User Interface for Building and Debugging Multi-Agent Systems (March 7, 2024, 3.07k views, 9 likes)
  • Getting Started with AutoGen - A Framework for Building Multi-Agent Generative AI Applications (February 5, 2024, 7.65k views, 36 likes)

I see a clear trend in the newsletter statistics - practical, getting-started content about AutoGen resonated strongly with readers, with the introductory guide receiving over 7,600 views and 36 likes. This aligns with the growing interest in multi-agent systems and validates the focus on educational content in this space.

The newsletter grew from 600 to 1600 subscribers this year! I am really grateful for the support and feedback from readers.

Community/Talks and Open Source

I gave a talk at QCon San Francisco "10 Reasons Your Multi-Agent Workflows Fail and What You Can Do About It" and another at the IEEE Vis Conference (NLViz workshop). I also participated in a panel on the future of data visualization at the Center for Design at Northeastern University.

This year had relatively fewer talks (2) compared to last year (7). It has mostly been a year of building and writing. Next year I plan to do more talks and workshops.

AutoGen and AutoGen Studio

Speaking of building, I am really excited about the progress being made with AutoGen . The new AutoGen v0.4 api which I have been contributing to introduces an async-first design, event-driven communication, and support for distributed implementations. This has been a significant step forward in making multi-agent systems more reliable and production-ready.

I also completely rewrote AutoGen Studio based on the new API, with a few new capabilities.

• Drag and drop declarative specification of multiagent teams

• Streaming agent updates

• Ability to interrupt task runs

• Gallery to import multi-agent team workflows

• And many more features coming.

P.S. For context - the original version of AutoGen Studio was written in January 2024, announced on the Microsoft Blog in June , and completely rewritten in November. It's been an intense year!"

Learning New Things

• DaVinci Resolve for video editing

Created my first technical tutorial video on AutoGen Studio using Davinci resolve. I learned quite a bit about the basics of a home recording setup - camera selection, lighting, audio etc. I learned about the basics of Davinci Resolve - audio effects, trimming, etc. I have tons to learn here, but it was a good start.

• Async Programming in Python

While I have general familiarity with async programming, this year I got to dive deeper into the details and implementation, especially in python - async/await, the asyncio library and related distributed programming patterns in Python. This has been useful as part of contributions to the new AutoGen Studio rewrite. On this front, I have learned a lot from colleagues - Jack Gerrits, Eric Zhu through numerous library design discussions, and code reviews for AutoGen.

• ReactFlow

I started exploring ReactFlow last year, but really dug in this year. Specifically, I have used it in implementing the drag and drop multi-agent composition view in AutoGen Studio. Some of the interesting areas include building custom nodes, layouting libraries, custom interactions etc.

• Languages - 312 Day Stream on Duolingo

Earlier this year, I figured it might be neat to try to learn a new language. So far I have a 312 day streak on Duolingo learning Russian. Its mostly doing 5 - 10 minute exercises (usually not more) each day. I suspect I need to keep at it for a few more years to get to a decent level. It feels great knowing I kept at it for this long.

Work-Life Balance

I didnt do as well as I would have liked on the fitness front. Injury recovery took a toll.

However, this year was rich in family time, which meant a lot to me. I got to spend many weekend hours with my son - mostly playing soccer together and just general calisthenics (monkey bars, pull ups, etc). It's been rewarding to see his progress - he can now do 10 pullups and 20 pushups from pretty much zero earlier in the year. We did some family travel - 3 international trips across the year.

Yes, organizing with sub-bullets would make the goals clearer:

Goals for 2025

For 2025, I'd like to accomplish the following:

• Book:

  • Complete the remaining 3 chapters of Multi-Agent Systems with AutoGen.
  • Add a use case gallery to the MultiAgent book website with code samples across industry verticals

• Writing:

  • Publish 10 long-form technical newsletter posts
  • Grow the newsletter to 4,000 subscribers
  • [Bonus / Stretch] Create a focused short course on getting started with multi-agent systems. I have gotten feedback from readers that this would be useful.

• Research:

  • Contribute to 5 research papers focusing on multi-agent systems and human-AI interaction

• Community:

  • Deliver at least 3 technical talks at key conferences
  • Continue contributions to AutoGen. Implement a new flow interaction in AutoGen Studio.

• Health and Work-Life Balance:

  • Work towards 3 workout sessions per week, with progress data
  • Continue to protect regular family time

Conclusion

2024 was a year that reminded me of the importance of resilience and adaptability. Alot of what is described above was done while severely disabled (the world suddenly feels very different when you have only one functional hand), that part was hard. Despite the challenges, it brought significant professional growth and meaningful family experiences. I'm grateful for the support from family, colleagues, and Microsoft!

Thank you for reading, and here's to an impactful 2025!