The Future of AI: Customizing AI agents with the Semantic Kernel agent framework

Microsoft

Mar 12, 2025

The Future of AI blog series is an evolving collection of posts from the AI Futures team in collaboration with subject matter experts across Microsoft. In this series, we explore tools and technologies that will drive the next generation of AI. Explore more at: https://aka.ms/the-future-of-ai

Customizing AI agents with the Semantic Kernel agent framework

AI agents are autonomous entities designed to solve complex tasks for humans. Compared to traditional software agents, AI-powered agents allow for more robust solutions with less coding. Individual AI agents have shown significant capabilities, achieving results previously not possible. The potential of these agents is enhanced when multiple specialized agents collaborate within a multi-agent system. Research has shown that such systems, comprising single-purpose agents, are more effective than single multi-purpose agents in many tasks [1]. This enables automation of more complex workflows with improved results and higher efficiency in the future.

In this post, we are going to explore how you can build single agents and multi-agent systems with Semantic Kernel.

Semantic Kernel is a lightweight and open-source SDK developed by Microsoft, designed to facilitate the creation of production-ready AI solutions. Despite its capabilities, Semantic Kernel remains accessible, allowing developers to start with minimal code. For scalable deployment, it offers advanced features such as telemetry, hooks, and filters to ensure the delivery of secure and responsible AI solutions.

The Semantic Kernel Agent Framework offers pro-code orchestration within the Semantic Kernel ecosystem, facilitating the development of AI agents and agentic patterns capable of addressing more complex tasks autonomously.

Starting with individual agents is recommended. Semantic Kernel provides a variety of AI service connectors, allowing developers and companies to select models from different providers or even local models. Additionally, Semantic Kernel gives developers the flexibility to integrate their agents created from managed services like Azure OpenAI Service Assistant API and Azure AI Agent Service into a unified system. Refer to the samples in the Semantic Kernel GitHub repository to get you started.

Previous posts have thoroughly examined the principles of designing single agents and the effectiveness of multi-agent systems. The objective of this post is not to determine when a single agent should be employed versus a multi-agent system; however, it is important to emphasize that agents should be designed with a single purpose to maximize their performance. Assigning multiple responsibilities or capabilities to a single agent is likely to result in suboptimal outcomes.

If your tasks can be efficiently accomplished by a single agent, that’s great! If you find that the performance of a single agent is unsatisfactory, you might consider employing multiple agents to collaboratively address your tasks. Our recent Microsoft Mechanics video outlines how a multi-agent system operates.

Semantic Kernel offers a highly configurable chat-based agentic pattern, with additional patterns coming soon. It accommodates two or more agents and supports custom strategies to manage the flow of chat, enhancing the system’s dynamism and overall intelligence.

Semantic Kernel is production-ready with built-in features that are off by default but available when needed. One such feature is observability. Often in an agentic application, agent interactions were not shown in the output, which is typical since users often focus on results. Nonetheless, being able to inspect the inner process is crucial to developers. Tracking interactions becomes challenging as the number of agents increases and tasks grow complex. Semantic Kernel can optionally emit telemetry data to ease debugging.

For a demonstration of three agents collaborating in real-time and reviewing the agent interactions with the tracing UI in Azure AI Foundry portal, please watch the following video demo:

The code to the demo can be found in a single demo app in the Semantic Kernel repository: semantic-kernel/python/samples/demos/document_generator at main · microsoft/semantic-kernel

In summary, Semantic Kernel offers an efficient framework for both single and multi-agent systems. As the platform evolves, it promises even more innovative patterns and capabilities, solidifying its role in agent-based AI. Whether for simple tasks or complex projects, Semantic Kernel provides the necessary tools to achieve your goals effectively. Happy coding!

To get started,