Lots of stuff happened this year - I finished my book (not without challenges), contributed the new Microsoft Agent Framework, and wrote more than I ever have.

My objectives for 2025, as outlined in my 2024 year-in-review post, included completing the book, growing the newsletter, and delivering more talks. Progress was made on all fronts. Work-life balance was mixed - strong start, then things got busy.

I still wrote quite a bit of code - but code metrics have increasingly less signal (beyond worked on lots of software!) given how much of this is enabled by coding agents.

The Book

The biggest milestone: finishing the book.

It wasn't without challenges - the project went through some twists along the way. But I shipped Designing Multi-Agent Systems in November: 15 chapters, 395 pages, 186 code snippets. It hit #1 New Release in Generative AI on Amazon and has sold ~500 copies so far (roughly half digital, half print) in about 5 weeks.

I'm also proud of building out the entire digital distribution platform at buy.multiagentbook.com - payment processing, file delivery, the works. Learned a lot.

Research, Work and OSS

Over the last few years, some of you have followed my work contributing to AutoGen and building AutoGen Studio. This year, the AutoGen and Semantic Kernel teams came together to create Microsoft Agent Framework, and a big chunk of my efforts went into this (together with a few re-orgs). Agent Framework provides a stronger developer experience and better integration with Microsoft Foundry Agent Service.

In addition to core framework contributions, I built DevUI - a UI tool for testing agents and workflows. It provides a chat interface, thread management, and event tracing that shows tool calls and execution flow. DevUI was showcased in multiple Ignite 2025 sessions (LAB513, BRK197) and featured in the Agent Framework launch video.

On the research side, I contributed to a few papers:

• AI Fatigue: Reflections on the Human Side of AI's Rapid Advancement — Victor Dibia (Communications of the ACM)

• Interactive Debugging and Steering of Multi-Agent AI Systems — Will Epperson, Gagan Bansal, Victor Dibia, Adam Fourney, Jack Gerrits, Erkang Zhu, Saleema Amershi (CHI 2025)

• Magentic-UI: Towards Human-in-the-loop Agentic Systems — Hussein Mozannar, Gagan Bansal, Cheng Tan, Adam Fourney, Victor Dibia, and others (arXiv preprint, CHI 2026 submission)

Newsletter, YouTube and Writing

The newsletter had its best year:

• 1,600 → 8,800+ subscribers (I had hoped for about 4,000 based on 2024 plans!)
• 29 posts published

Key themes: agent framework comparisons, MCP deep dives, UX for autonomous agents, and agent fundamentals. I also reflected on why I write about AI and the cognitive costs of AI coding tools. Clearly, MCP did seem to get most attention (views) of all I wrote about this year.

I also published 16 YouTube tutorials on my Designing With ML channel, totaling ~100k views. The videos covered AutoGen, MCP, and agent frameworks - mostly companion content to the newsletter posts. Teaching through video is harder than writing, but it reaches a different audience and forces clearer explanations.

Talks

• AI.Engineer World's Fair — Keynote on UX design principles for multi-agent systems
• MCP Dev Days — "Agents Talking to Agents on MCP"
• Northeastern University - Introduction to Multi-Agent Systems and Applications in Visualization (guest lecture, with Prof Enrico Bertini)

Work-Life Balance

Mixed results this year.

The first three months were strong - consistent with diet, fitting in 3-5 workouts per week, lost about 8kg. Then things got busy at work (including the AutoGen + Semantic Kernel merge) and the push to finalize the book. Fitness fell off (to mostly the outdoor ~1.5hrs on the weekend with my 6yr old either playing soccer or basketball).

But I did spend a lot of quality time with my son. This year he got interested in basketball - something I hadn't played seriously in years. We've averaged about 1.5 - 2 hours of play most weekends over the last few months. It's been great watching him learn and having a reason to get back on the court myself.

Also managed three family trips this year.

Learning New Things

Learning outside of work has been a feature of my year-in-review reflections. Last year I studied Russian on Duolingo for 312 days and started learning DaVinci Resolve for video editing. This year, the biggest non-work learning came from the challenges of self-publishing my book.

• Cover Design: Learned the details of bleeds, color mapping, CMYK color space for print, and exporting to formats required by print services. Did it all myself.

• Typesetting with Quarto: Dug deep into the Quarto publishing system. My research background with LaTeX typesetting helped here. Researched and implemented the entire self-publishing pipeline - from manuscript to distribution.

• Marketing and Visibility: Spent some time revisiting how advertising strategies work - bids, keyword selection, data analysis on campaigns and how they feed into the design of new campaigns. Got familiar with terms like ACOS, revisted SEO optimization, and how that space all works.

• Digital Sales Platform: Built and optimized the PDF sales platform at buy.multiagentbook.com - clean UX from payment to email delivery with lifetime access, all optimized to run < $2/month in infrastructure costs.

It was long hours (especially between July and November), but tons of learning and fun.

What Didn't Work

A year-in-review isn't complete without acknowledging what didn't go as planned. My 2024 goals included optimistic targets of 5 research papers and 3 tracked workouts per week. I hit neither.

But the reasons matter:

• Exercise: This was a tradeoff. I managed 3+ workouts per week for the first 5 months, but that fell off once really buckled in to finish the book. June through November, every free moment went into finishing the book. Fun fact, I wrote huge chunks of the book while sitting either at the back of my car or a cafe on the weekends while waiting for my 6yr old to finish some of this activities (gynmastics, scout). I think this is why most authors strongly recommend against writing books :).

• Research papers: I got re-orged from Microsoft Research to the Azure Foundry product team - research and paper-writing just aren't part of the job focus anymore. The goal shifted to enabling successful tech transfer for AutoGen (which I'm proud of) and shipping enterprise features (I contributed to the multi-agent workflows feature - initial prototypes, UX reviews, etc.).

What I've learned: goals exist in context, and context changes. What matters is adapting. I recommend this useful article on rituals over resolutions.

Goals for 2026

Nothing too ambitious - just continue learning new things, protect family time, and stay healthy. Professionally, I'd like to spend more time training sub-8B parameter models for specific domains (this will likely be important going forward) and build a portfolio of projects in that area.

• Writing: Publish at least 15 newsletter posts this year
• Newsletter: Continu to write and teach on the newsletter. Ideally, the work continues to inspire and newsletter grows - perhaps to 15,000 subscribers (a stretch goal, but feeling optimistic given the unexpected growth this past year!)
• Model Training: Train and release at least 2 models on Hugging Face for domain-specific areas - e.g., visualization or agents
• Research: Write and submit 2 research papers - I envision these will be more technical reports detailing model or engineering work.
• Career: During this year, I worked on several projects that cut across multiple teams and orgs - the work on integrating agent framework meant working with core Azure product folks, desingners, PMs, setting technical direction, providing reviews with feedback. I'd like to continue to learn and grow in this area.

Conclusion

Grateful for colleagues, the 8,800+ newsletter subscribers, and everyone who picked up the book.

Here's to 2026.