At the Nashua Cloud .NET User Group, I had the opportunity to present on AI Workflows with Microsoft Agent Framework.
The main idea behind the session was simple: as AI becomes part of real business systems, we need more than just prompts and agent loops. We need structure, control, observability, and predictable execution.
Microsoft Agent Framework gives developers a way to build AI agents and orchestrate them through workflows using .NET or Python. Instead of relying on an agent to decide every step dynamically, workflows allow us to define clear execution paths using concepts like executors, edges, workflow graphs, and typed messages.
During the session, we looked at the difference between agents and workflows. Agents are useful when we want LLM-driven reasoning, but workflows are important when we need predefined paths, deterministic behavior, compliance, and enterprise governance. In production systems, that distinction matters.
We walked through several workflow patterns, including sequential workflows, concurrent workflows, group chat workflows, and human-in-the-loop scenarios. Each pattern solves a different kind of problem. Sequential workflows are great when steps need to happen in order. Concurrent workflows help when multiple tasks can run in parallel. Group chat workflows allow multiple agents to collaborate, while human-in-the-loop workflows bring approval, review, or exception handling into the process.
One of the key takeaways for me is that workflows make AI systems easier to reason about. They give architects and developers a clearer way to design systems that are testable, observable, and governable. This becomes especially important when AI is used in enterprise scenarios where reliability and accountability matter.
We also discussed how workflows and Agent-to-Agent communication can work together. A2A helps agents talk across runtimes or organizations, while workflows help choreograph how those agents participate in a larger process. In other words, A2A is how agents communicate; workflows are how we design the overall execution.
The session ended with demos around deterministic workflow orchestration, including sequence, concurrent execution, group chat, and human-in-the-loop flows.
Thanks to everyone who joined the session and continues to support the Nashua Cloud .NET User Group. These community conversations are where we get to explore not just what is new, but what is practical, useful, and ready for real-world systems.
Presentation: https://www.slideshare.net/slideshow/microsoft-agent-framework-ai-workflows/287480032
Sample Code: https://github.com/nhcloud/ai-workflow
If you missed the session, join us at the next Nashua Cloud .NET User Group (NashuaUG) meetup to continue exploring practical, real-world AI engineering.