Semantic Kernel

The kernel holds AI services and plugins; an Agent wraps a service, instructions and plugins, and an AgentThread holds conversation state (some agent types require a matching server-side thread, such as AzureAIAgentThread). Tools are kernel functions, which can come from code, prompt templates, OpenAPI specs or MCP servers.

Multi-agent coordination is offered as orchestration classes (concurrent, sequential, handoff, group chat with a manager, and Magentic) that share one construction and invoke pattern and run on a runtime such as InProcessRuntime. Agents can also be passed to another agent as plugins, as in the README triage example. The docs mark orchestration as experimental and not yet available in Java.

Longer workflows can use the experimental Process Framework, where steps react to events. Memory beyond the thread comes from providers such as Mem0 (experimental) and from vector store connectors.

Handoff and group chat orchestrations accept a human-response callback (human_response_function in Python, InteractiveCallback in .NET) that is called whenever an agent needs user input; response callbacks let an app watch each agent message. In the Process Framework, a step can wait for an external approval event before continuing, as shown in the human-in-the-loop example. Function invocation filters can inspect calls and terminate auto function calling.

pip install semantic-kernel

import asyncio
from semantic_kernel.agents import ChatCompletionAgent, SequentialOrchestration
from semantic_kernel.agents.runtime import InProcessRuntime
from semantic_kernel.connectors.ai.open_ai import OpenAIChatCompletion

async def main():
    service = OpenAIChatCompletion()  # reads OPENAI_API_KEY and OPENAI_CHAT_MODEL_ID
    outliner = ChatCompletionAgent(service=service, name="Outliner", instructions="Turn the request into a 3-point outline.")
    writer = ChatCompletionAgent(service=service, name="Writer", instructions="Write one short paragraph per outline point.")

    orchestration = SequentialOrchestration(members=[outliner, writer])
    runtime = InProcessRuntime()
    runtime.start()
    result = await orchestration.invoke(task="Explain what a vector index is", runtime=runtime)
    print(await result.get(timeout=60))
    await runtime.stop_when_idle()

asyncio.run(main())

• Requirements per the README: Python 3.10+, .NET 10.0+, or JDK 17+.
• Set OPENAI_API_KEY (or AZURE_OPENAI_API_KEY for Azure) before running; the OpenAI connector also reads OPENAI_CHAT_MODEL_ID.
• Agent orchestration, agent memory and the Process Framework are experimental and may change; orchestration is not available in the Java SDK.
• .NET packages are Microsoft.SemanticKernel plus Microsoft.SemanticKernel.Agents.Core; Java lives in the separate semantic-kernel-java repository.
• Microsoft publishes a migration guide from Semantic Kernel to Microsoft Agent Framework.

Microsoft's README says Semantic Kernel is now Microsoft Agent Framework and calls the framework its successor. Microsoft's blog says Semantic Kernel 1.x keeps getting critical bug and security fixes, with most new features going to Agent Framework.

Yes. The Python SDK loads MCP servers as plugins (stdio, SSE and Streamable HTTP); .NET has repository samples, and the docs mark .NET and Java MCP docs as coming soon.

Partly. The .NET SDK includes A2AAgent and A2AHostAgent for calling and hosting A2A agents; no Python or Java A2A support was found.

C#/.NET and Python in the main repository, and Java in the separate semantic-kernel-java repository. Agent orchestration is not yet available in Java.

Yes. It is MIT-licensed and has no paid tier; you pay for whichever model service you connect.