Gagan Bansal

Gagan Bansal

Principal Researcher, Microsoft Research

I build AI agent systems and study how to keep humans in control of them.

AutoGen Co-lead

Multi-agent framework that is now the core of Microsoft Agent Framework.

View

MarkItDown Co-lead

Convert any file to Markdown for LLM pipelines.

View

Magentic-One Co-lead

State-of-the-art multi-agent system for web and file tasks.

View

Magentic-UI Co-lead

Human-centered web agent with co-planning and guardrails.

View

Magentic Marketplace Co-lead

Simulation environment for studying AI-powered two-sided markets.

View

Selected papers. Full list on Google Scholar

2025
Magentic-UI: Human-in-the-loop Agentic Systems Tech Report Web agents need human oversight. We built one that plans with users and asks before acting.
2025
Magentic Marketplace: Studying Agentic Markets Tech Report When AI agents buy and sell on our behalf, markets change. We built a simulation to study how.
2025
Challenges in Human-Agent Communication Tech Report Agents fail in ways chatbots don't. We identify the communication breakdowns and propose fixes.
2025
Generation Probabilities Are Not Enough: Uncertainty in Code Completions ToCHI Model confidence scores don't help programmers spot bad suggestions. Highlighting uncertain tokens does.
2024
AutoGen: Multi-Agent Conversation Framework COLM Best Paper, ICLR Workshop LLMs work better in teams. We built a framework that lets multiple agents collaborate through conversation.
2024
Magentic-One: A Generalist Multi-Agent System Tech Report One agent struggles with complex tasks. Five specialized agents, orchestrated together, achieve state-of-the-art.
2024
Reading Between the Lines: AI-Assisted Programming CHI Honorable Mention We measured how programmers actually use Copilot. A large amount of time goes to verifying suggestions, not writing code. Check out the visuals based on real usage data.
2021
Does the Whole Exceed its Parts? AI Explanations and Team Performance CHI AI explanations help, but not always. They work when humans can catch AI mistakes, not just agree with them.
AutoGen v0.4 AutoGen v0.4: A New Architecture for Agentic AI Challenges in Human-Agent Communication Challenges in Human-Agent Communication Magentic Marketplace Magentic Marketplace: Studying Agentic Markets Does the Whole Exceed its Parts? Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance

Now

Principal Researcher at Microsoft Research AI Frontiers, where I conduct interdisciplinary research bridging artificial intelligence and human-computer interaction. I am one of the research leads of AutoGen, an open-source multi-agent framework that became the foundation of Microsoft's Agent Framework in 2025. I also co-lead Magentic-One (a generalist multi-agent system), Magentic-UI (its human-centered interface), MarkItDown (document-to-markdown conversion), and Magentic-Marketplace (for studying agent economies).

Research

My research focuses on keeping humans in control of AI agents. This includes studying how agents fail, how they might manipulate users, and how to design effective oversight mechanisms. Before agents, I worked on human-AI decision making—specifically how AI explanations affect team performance and how to update AI systems without breaking user trust.

Before

Ph.D. in Computer Science from University of Washington, advised by Dan Weld. During my PhD, I interned at Microsoft Research with Besmira Nushi, Ece Kamar, and Eric Horvitz. B.Tech from IIT Delhi, where I worked with Mausam on natural language processing.