Microsoft Agent Framework

1. Project Overview

The project comes with key diagrams:

! Microsoft Agent Framework Banner

Video entry in the official README:

! Microsoft Agent Framework Introduction

2. What does it mostly do?

The Microsoft Agent Framework provides two broad categories of capabilities:

Compared with the common Agent SDK, MAF emphasizes from prototype to production:

3. Relationship with AutoGen / Semantic Kernel

A large part of MAF's pre-sale value comes from the "official succession route".

The MS Learn overview makes it clear that the Agent Framework combines AutoGen simple Agent abstraction with enterprise-level Kernel-Semantic capabilities, such as session-based state management, type safety, middleware/filters, telemetry, and richer model and embedding support. It also adds a graph-based workflows for explicit multi-Agent execution path control.

Pre-sales words:

"If the customer is still evaluating the AutoGen, it is time to directly include the Microsoft Agent Framework as a primary option, as it is Microsoft current new route to production-level Agents."

4. Core Competence List

5. What about Agents vs Workflows

MS Learn gives a good judgment standard:

When using agent... | Use Workflow when... |

|---|---|

| The task is open and conversational | The process has clear steps |

| Requires autonomous tool use and planning | Requires explicit control over execution order |

| A single LLM call or a tool call is enough. | Multiple agents or functions must work together. |

There is another sentence that is very useful before sales:

If you can use ordinary functions to handle tasks, use functions instead of AI for AI.

The pre-sales value behind this sentence is that MAF does not encourage turning everything into an Agent, but puts the Agent into the links that really require reasoning, tool invocation, and human-computer collaboration; The deterministic steps are still controlled by ordinary code/Workflow. This is more in line with the requirements of enterprise customers for stability, cost and controllability.

6. Architecture/deployment/integration approach

The typical architecture of MAF can be understood:

1. Provider layer: Connect Microsoft Foundry, Azure OpenAI, OpenAI, other model vendors.
2. Agent layer: build Agent with model, instructions, tools and session.
3. Tool/MCP layer: connect enterprise API, MCP Server, hosted tools, code interpreter, file retrieval, etc.
4. Workflow layer: schema arrangement Agent, function, manual approval, checkpoint, branch routing.
5. Observability/Governance layer: OpenTelemetry, middleware, log, policy, content filtering.
6. Hosting tier: local development, Azure Functions, Durable Agents/Workflows, Foundry Hosted Agents, A2A.

When targeting Azure customers, focus on:

How to use #7.

Python installation:

pip install agent-framework
pip install azure-identity

. NET installation:

dotnet add package Microsoft.Agents.AI
dotnet add package Microsoft.Agents.AI.Foundry
dotnet add package Azure.AI.Projects
dotnet add package Azure.Identity

Python Minimal Example:

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

async def main():
    client = FoundryChatClient(
        credential=AzureCliCredential(),
        # project_endpoint="https://your-project.services.ai.azure.com",
        # model="gpt-4o",
    )

    agent = Agent(
        client=client,
        name="HelloAgent",
        instructions="You are a friendly assistant. Keep your answers brief.",
    )

    result = await agent.run("What is the capital of France?")
    print(result)

if __name__ == "__main__":
    asyncio.run(main())

Streaming output:

async for chunk in agent.run("Tell me a one-sentence fun fact.", stream=True):
    if chunk.text:
        print(chunk.text, end="", flush=True)

Note: MS Learn explicitly reminds the Agent that the '.env' file Framework not be loaded automatically; if'.env' is required, load_dotenv()'is explicitly called in the application'.

8. Applicable Scenario

9. Not quite the scene

10. What can I say before sales

One-sentence positioning:

"Microsoft Agent Framework is a new generation framework for Microsoft current production-level Agent and multi-Agent workflows, unifying AutoGen multi-Agent thinking with Kernel-Semantic enterprise capabilities, and is suitable for moving from prototype to governed, observable, and recoverable enterprise-level Agent applications."

Customer Value:

11. Comparison with AutoGen / CrewAI / LangGraph

Pre-sales advice:

1. If the customer is an Azure/Microsoft ecosystem, MAF should have priority access to the scenario.
2. If the customer already has a AutoGen,MAF is the main migration line.
3. If the customer is a pure Python, non-Azure, partial business automation demo, you can compare CrewAI at the same time.
4. If the customer already has a LangChain/LangGraph ecosystem, it needs to be selected from model Provider, state management, workflow governance, and deployment requirements.

12. PoC Recommendations

MAF PoC does not recommend only "Hello Agent", but should choose a process that can reflect the production capacity.

Suggested indicators:

13. Risks and Considerations

1. The GitHub API is anonymously throttling this time: The GitHub API failed to verify the stars, forks, and latest GitHub release again. The package version in the note uses PyPI/NuGet to query the results.
2. Third-party system risk: README Important Note clearly states that third-party servers, agents, code, and non-Azure Direct models are all subject to customer risk, review data flow and cost.
3. Be cautious with Azure certification: MS Learn reminds that 'DefaultAzureCredential' is easy to develop, but production recommends using more explicit credential, such as Managed Identity.
4. Provider capabilities are not completely equivalent: some provider/capabilities in the AutoGen migration guide are still planned and need to be confirmed item by item.
5. SDK is not a complete platform: MAF provides framework and samples, and production still requires application layer permissions, auditing, UI, CI/CD, deployment and monitoring.
6. AI the boundary of responsibility: the customer needs to implement Responsible AI mitigation measures such as metaprompt, content filters, safety evaluation, quality control, etc.
7. The framework is still evolving rapidly: PyPI/NuGet versions are updated faster, and production PoC requires fixed versions and upgrade policies.

14. Frequently Asked Customer Questions

15. My Pre-Sales Judgment

Microsoft Agent Framework is the most important Agent framework in the current Microsoft ecology. Its positioning is very clear: not to make another Agent demo framework, but to make infrastructure for enterprises to push Agent from demo to production. It combines AutoGen multi-Agent thinking, Kernel-Semantic enterprise engineering capabilities, and Azure/Foundry deployment ecology into one line.

If the customer is already in the Microsoft ecosystem, MAF is a very natural recommended path. Especially when customers mention these keywords: Azure OpenAI, Microsoft Foundry,. NET, AutoGen migration, production-level workflow, observable, human review, long task recovery, and corporate governance can all put MAF at the core of the program.

However, if the customer just wants to be a simple chat robot quickly, MAF may be biased. If the customer does not touch the Microsoft ecology at all, he should also make a horizontal comparison with CrewAI, LangGraph, etc. When it comes to pre-sales promotion, the best entry point is not "it can write a poem", but "it can put the Agent into a recoverable, auditable, and human-intervention business process".

16. REFERENCE

-GitHub: https://github.com/microsoft/agent-framework

-MS Learn Documentation: https://learn.microsoft.com/en-us/agent-framework/

-Framework Overview: https://learn.microsoft.com/en-us/agent-framework/overview/agent-framework-overview

-Quick Start: https://learn.microsoft.com/en-us/agent-framework/tutorials/quick-start

-AutoGen Migration Guide: https://learn.microsoft.com/en-us/agent-framework/migration-guide/from-autogen/

-Semantic Kernel Migration Guide: https://learn.microsoft.com/en-us/agent-framework/migration-guide/from-semantic-kernel

-PyPI: https://pypi.org/project/agent-framework/

-NuGet: https://www.nuget.org/profiles/MicrosoftAgentFramework/