markus41/plugins/microsoft-agents-expert/skills/agent-framework

Microsoft Agent Framework

Open-source, multi-language SDK for AI agents and graph-based workflows — the direct successor to both Semantic Kernel and AutoGen, built by the same teams. AutoGen's simple agent abstractions + SK's enterprise features (sessions, type safety, middleware, telemetry), plus typed workflows. Docs: https://learn.microsoft.com/agent-framework/overview/ · Repo: github.com/microsoft/agent-framework (Go: agent-framework-go).

Status (mid-2026): C#/.NET core is stable v1.x; Python is primary alongside it. Many provider/integration packages are still preview (Mem0/Redis/Neo4j providers, AG-UI, Dev UI). Go is public preview. Check package status before promising GA to a user.

Packages

| Language | Install | |---|---| | .NET | Microsoft.Agents.AI (core), .Abstractions, .OpenAI, .Foundry, .A2A (NuGet) | | Python | pip install agent-framework (meta) → agent-framework-core + providers: agent-framework-foundry, agent-framework-openai, agent-framework-copilotstudio, agent-framework-mem0, agent-framework-foundry-hosting | | Go | go get github.com/microsoft/agent-framework-go (preview) |

Python imports come from agent_framework (e.g. from agent_framework.foundry import FoundryChatClient).

Core abstractions

• .NET: AIAgent (base), ChatClientAgent (wraps any Microsoft.Extensions.AI.IChatClient), AgentSession for multi-turn state, RunAsync / RunAsync<T> (structured output → AgentResponse<T>).
• Python: Agent, BaseAgent, SupportsAgentRun, AgentSession, @tool decorator.
• Special agents: CopilotStudioAgent (call a Copilot Studio agent from code), A2AAgent (remote agent over the A2A protocol).
• Three pillars: Agents (single agent loop), Harness (opinionated long-task agent: planning/todo tracking, context compaction, file access/memory, don't-ask-again tool approval), Workflows (typed graphs).

from agent_framework import Agent
from agent_framework.foundry import FoundryChatClient
from azure.identity import AzureCliCredential

agent = Agent(
    client=FoundryChatClient(credential=AzureCliCredential()),
    instructions="You are a helpful assistant",
)
result = await agent.run("Help me with this task")

using Azure.AI.Projects;
using Azure.Identity;
using Microsoft.Agents.AI;

AIAgent agent = new AIProjectClient(new Uri("<project-endpoint>"), new DefaultAzureCredential())
    .AsAIAgent(model: "gpt-4o-mini", name: "Joker", instructions: "You are good at telling jokes.");
AgentSession session = await agent.CreateSessionAsync();
Console.WriteLine(await agent.RunAsync("Tell me a joke.", session));

Workflows and orchestration

Typed graphs of executors connected by edges (direct, conditional, switch-case, fan-out, fan-in), with checkpointing and human-in-the-loop pauses. High-level builders in agent_framework.orchestrations / the .NET equivalents:

| Pattern | Builder | Use when | |---|---|---| | Sequential | SequentialBuilder | Pipeline: research → write → review | | Concurrent | fan-out/fan-in edges | Independent subtasks in parallel | | Handoff | HandoffBuilder | Triage agent routes to specialists | | Group Chat | GroupChatBuilder | Agents debate/collaborate on shared thread | | Magentic | Magentic orchestration | Manager dynamically plans and coordinates specialists |

A workflow can itself be exposed as an agent: workflow.as_agent(name="Content Pipeline Agent").

from agent_framework.orchestrations import SequentialBuilder
workflow = SequentialBuilder(participants=[researcher, writer, reviewer]).build()

Middleware, memory, tools

• Middleware (function- or class-based) intercepts agent actions — logging, guardrails, caching, tool-call gating. Prefer middleware over prompt hacks for policy enforcement.
• Memory: context providers (chat-history memory is released; Mem0/Redis/Neo4j are preview) and chat-history providers (Cosmos DB, Redis) for persistence.
• Tools: native functions (@tool / AIFunctionFactory), plus MCP clients — MCPStreamableHTTPTool, MCPStdioTool, MCPWebsocketTool — and hosted MCP via get_mcp_tool(...) on Foundry/OpenAI/Anthropic chat clients.

Observability

OpenTelemetry per the GenAI semantic conventions — configure_otel_providers() in Python; exports to Azure Monitor or the Aspire Dashboard. MCP trace context propagates automatically via _meta. Instrument from day one; agent bugs are trace-shaped.

Practices

• Wrap any IChatClient-compatible model — the framework is provider-agnostic; don't hard-code a single vendor path when the user needs flexibility.
• Use AgentSession for multi-turn state instead of hand-rolled history lists.
• Structured outputs via RunAsync<T> beat string parsing for machine-consumed results.
• Migrating from Semantic Kernel or AutoGen? Agent Framework is the successor — map SK planners/agents and AutoGen group chats onto the orchestration builders above (see /msagent-migrate).
• Deploy targets: self-host (any ASP.NET/ASGI app), inside the M365 Agents SDK for channel reach (skills/m365-agents-sdk), or as a Foundry hosted agent (agent-framework-foundry-hosting, ResponsesHostServer — see skills/microsoft-foundry).