AI Agent Architect

Victor Dibia Portfolio

Principal Research Software Engineer at Microsoft

Principal Research Software Engineer at Microsoft and core contributor to AutoGen and its successor, the Microsoft Agent Framework; author of the 2025 book "Designing Multi-Agent Systems," which walks through agent design patterns and implementation with 46 diagrams and 154 code examples, and built AutoGen Studio, a no-code multi-agent workflow builder.

Multi-Agent SystemsAutoGenAgent OrchestrationMicrosoft

Location

Coordinates unknown. Vibes: international.

Companies

Microsoft

Skills

Multi-Agent SystemsAutoGenAgent OrchestrationMicrosoft

What makes it work

A breakdown of the choices that make this portfolio stand out.

1

A full book on agent design patterns, not just a framework README

Designing Multi-Agent Systems formalizes patterns (coordination, role assignment, communication) into a structured reference, a more durable teaching artifact than scattered blog posts about the same framework.

2

Both the low-level framework and a no-code layer on top of it

Building AutoGen (the programmable core) and AutoGen Studio (a visual interface on top of it) shows a design sense that spans from framework internals to how less technical users actually construct agents.

3

A body of published, peer-reviewed research underneath the applied tools

Citations in ACL, EMNLP, AAAI, and CHI venues alongside patents show the applied agent-framework work is backed by a real academic research record, not framework engineering alone.

What AI Agent Architects can take from this

Specific, actionable tips to apply to your own portfolio — no generic advice.

  • If you understand the recurring design patterns in your framework, write them up as a structured reference, not just scattered documentation.

  • Build both the low-level tool and a simplified interface on top of it — it shows you understand the pattern deeply enough to make it accessible.

  • Where your applied work has research underneath it, cite the papers directly — it distinguishes engineering built on studied principles from engineering by trial and error.

Case studies

A book, a low-code tool and a library, all built around one question: how people and AI agents work together. Each one comes with code you can run, including the from-scratch agent library the book builds chapter by chapter.

Case StudyadvancedPublished 2025; revised Aug 2026

Designing Multi-Agent Systems: Principles, Patterns and Implementation for AI Agents

PythonPicoAgentsFastAPIServer-Sent EventsPydanticBrowser automation

A 15-chapter book in four parts. It covers multi-agent concepts and orchestration patterns, then builds PicoAgents, a complete library with agents, tools, memory, workflows and orchestration, from scratch. Later parts cover evaluation, optimization and responsible AI.

Teaching agent design from first principles, not through one framework's API, shows he understands the layer below the tools. The code examples run, which keeps the book testable.

  • Taxonomy from deterministic workflows to autonomous orchestration

  • Builds a full agent library (PicoAgents) chapter by chapter

  • Chapters on computer-use agents, evaluation via trajectories, and UX for delegation

Projects

Open SourceadvancedPaper Aug 2024; ongoing in microsoft/autogen

AutoGen Studio

PythonAutoGenFastAPISQLModelReactJSON specs

A low-code web interface and Python API for prototyping, debugging and evaluating multi-agent workflows built on AutoGen. Agents are defined in a declarative JSON spec. The UI offers drag-and-drop workflow building, interactive debugging and a gallery of reusable components.

It brings developer-tool design to a hard problem: seeing and fixing what a group of agents did. He built it and wrote up the design reasoning.

  • Declarative JSON specification for agents, tools and workflows

  • Drag-and-drop builder with interactive run-and-debug

  • Four published design principles for no-code agent tools

Open SourceintermediatearXiv Mar 2023; ACL 2023 (Demo track)

LIDA

PythonLLMsImage generation modelsMatplotlibAltairSeaborn

A tool that generates visualizations and infographics from a dataset, with four modules. A summarizer turns data into a compact natural-language description, a goal explorer suggests what to chart, a VisGenerator writes, repairs and filters plotting code in any grammar, and an infographer styles the output with image models.

It is a multi-stage LLM pipeline in which generated code gets executed, checked and repaired, the same generate-and-verify loop used by today's coding agents.

  • Four-stage pipeline: summarize → set goals → generate code → stylize

  • Grammar-agnostic: works with any plotting library expressed as code

  • Self-evaluation and automatic repair of generated charts

Ready to build your portfolio?

Follow the AI Agent Architect roadmap — skills, projects, and timeline to get hired.

AI Agent Architect Roadmap