Victor Dibia is a principal research software engineer in Microsoft CoreAI, a founding contributor to AutoGen, and the creator of AutoGen Studio. He designs tools that help people build and evaluate AI agents while retaining visibility and control over their decisions.
Dibia holds a doctorate in information systems from City University of Hong Kong and a master’s degree in information networking from Carnegie Mellon University. He previously worked on human-computer interaction at IBM Research and machine learning at Cloudera before joining Microsoft Research’s Human-AI eXperiences team. Earlier projects included HandTrack.js, an open-source browser-based hand-detection library, and Data2Vis, which explored automatically generating data visualizations with neural networks.
At Microsoft, Dibia helped evaluate coding assistants and inform model selection for GitHub Copilot. He created LIDA, a system for model-assisted data visualization that summarizes datasets, proposes analytical questions, generates and executes visualization code, and produces charts or infographics. Published in 2023, the project anticipated agentic workflows organized around successive stages of planning, execution, and error recovery.
Dibia became a founding and core developer of AutoGen, contributed to its architectural rewrite and AgentChat interface, and created AutoGen Studio, a visual environment for assembling, testing, and debugging multi-agent applications. He also contributed to Magentic-One, Magentic-UI, and the foundational architecture of Microsoft Agent Framework, including its DevUI debugging environment. His more recent work includes Agent Optimizer, which evaluates agent configurations against explicit criteria, compares alternatives, and supports deploying stronger versions.
Principles behind his work
- Evaluate before adding agents. Define the task, establish a non-agent baseline, build dependable tools, and create task-specific tests before introducing autonomous behavior. Additional planners, verifiers, and collaborators should demonstrate measurable improvement.
- Match autonomy to the problem. Predetermined workflows suit tasks with known steps; exploratory problems may require agents that adapt as their actions change the environment. Multi-agent orchestration adds error surfaces and should be reserved for problems that genuinely benefit from planning, specialized perspectives, or extensive context.
- Make agent behavior inspectable and interruptible. Users need visible execution traces, provenance, resource consumption, and risk-aware human approval. Interruptibility, checkpointing, and rollback let them pause unsafe actions, restore earlier states, and resume reliably.
- Test agent design against consequential environments. His BlenderLM prototype translates natural-language requests into operations within Blender: a planner decomposes the task while a verifier checks scene state and visual output. The project makes changing environmental conditions, tool reliability, and human oversight concrete.
Dibia extends these ideas in Designing Multi-Agent Systems, accompanied by PicoAgents, an open-source implementation covering agent loops, orchestration, evaluation, observability, and human approval.
