Building Production AI Agents with Semantic Kernel, .NET Aspire, and Azure OpenAI
By Tomi Asp
Tomi Asp demonstrates how to build enterprise AI agents using Semantic Kernel, .NET Aspire, and Azure OpenAI — from plugins and tool calling to token management in production.
AI development has moved quickly from simple chatbots toward autonomous AI agents that orchestrate tasks, call tools, and make decisions. In this video, Mallow's Tomi Asp demonstrates how to build enterprise-grade agents using Microsoft's modern stack: Semantic Kernel and .NET Aspire.
What is Semantic Kernel?
Semantic Kernel is a lightweight, open-source SDK for integrating AI agents into C#, Python, or Java code. It acts as a flexible middleware layer that connects language models (like Azure OpenAI) to existing APIs and code.
Key benefits:
- Future-proof: Swap the underlying AI model without rewriting your application.
- Enterprise-ready: Built-in support for telemetry, security, and responsible AI controls.
- Modular: Expose existing code to the agent as plugins, allowing AI to execute real business processes.
What is .NET Aspire?
.NET Aspire streamlines building, running, and managing distributed applications. It provides a unified toolchain that eliminates complex configurations and makes local development and debugging straightforward.
Key features:
- Unified developer experience: Launch and test the entire distributed application and its dependencies with a single command.
- Code-first configuration: Architecture is defined in code, enabling type safety and version control for infrastructure.
- Orchestration and dashboards: A built-in dashboard provides direct visibility into service logging, telemetry, and performance.
AI agents in action
In the video, Tomi walks through a live demonstration of Semantic Kernel and .NET Aspire working together with AI agents. The demo follows an agent autonomously orchestrating tasks, calling tools, and communicating across different data sources.
A key takeaway: the pro-code approach gives developers significantly more control and ability to build valuable, real-data use cases compared to no-code tools.
Tokens and memory management: the real challenge
While the technical implementation with Semantic Kernel is straightforward, the biggest challenge in building agents is managing memory and token consumption. Every message grows the agent's conversation history, which can lead to exponentially rising costs. Tomi demonstrates how to compress history using the history reducer pattern.
This is where experience matters. Building a concept agent is easy — a production-ready solution requires careful architecture, cost management, and resilient MLOps practices.
Interested in building AI agents for your business? Get in touch — let's build together.
