Microsoft Agent Framework

What it does

Writes code

PartlyAn SDK rather than a coding product: the GitHub Copilot SDK backend supplies coding-oriented capabilities including file operations and shell execution, but what the agent writes is what you build.GitHub Copilot agents provide access to powerful coding-oriented AI capabilities, including shell command execution, file operations, URL fetching,…learn.microsoft.com/en-us/agent-framewo… ↗

Whole-codebase work

PartlyThe Harness targets long multi-step work with context compaction, file memory and todo tracking, and shell tools default to working-directory confinement; the vendor documents no repository indexing or repo-wide migration.A harness wraps a chat client with the scaffolding an agent needs to work through long, multi-step tasks - planning / execution modes, a todo list to…learn.microsoft.com/en-us/agent-framewo… ↗

Ships it

PartlyThe beta agent-framework-tools package gives an agent local or Docker shell execution, so it can run any command you approve; no deploy, release or pull-request step is documented.Shell execution can modify files, launch processes, access credentials, and communicate with external systems. Use the least-privileged execution…learn.microsoft.com/en-us/agent-framewo… ↗

Done-for-you service

NoYou run the software. Microsoft publishes an SDK and support is community channels: docs, GitHub issues, discussions and office hours.Microsoft Agent Framework is an open-source SDK for building AI agents that can reason, use tools, and interact with users and other agents.learn.microsoft.com/en-us/agent-framewo… ↗

How much it does unattended

Runs autonomously

YesThe Harness keeps an agent progressing through long multi-step tasks, with todo tracking and plan/execute modes on by default and optional bounded looping.It drives model and tool calls, manages conversation state and context, applies approval policies, and can keep the agent progressing through a…learn.microsoft.com/en-us/agent-framewo… ↗

Multi-agent

YesFive built-in orchestration patterns (sequential, concurrent, handoff, group chat and magentic manager), plus agents callable as tools and experimental background child agents.Agent Framework provides several built-in multi-agent orchestration patterns:learn.microsoft.com/en-us/agent-framewo… ↗

Agent permissions

YesApproval is per tool and declarative: LocalShellTool asks for every command and disabling that requires acknowledge_unsafe, ShellPolicy pre-filters commands, and the Copilot backend cannot touch shell, files or URLs without a permission handler.It defaults to a persistent shell, a 30-second timeout, 64-KiB output truncation, working-directory confinement, and approval for every command.learn.microsoft.com/en-us/agent-framewo… ↗

Which models it runs on

Claude

YesAnthropic is a first-class provider, reachable through Anthropic direct, Microsoft Foundry, Amazon Bedrock or Vertex AI, with function tools, code interpreter, web search and MCP tools supported.Anthropic — Claude inference through Anthropic, Foundry, Amazon Bedrock, or Vertex AI.learn.microsoft.com/en-us/agent-framewo… ↗

GPT

YesOpenAI and Azure OpenAI are supported through both Chat Completions and Responses clients; the provider table marks function tools, structured outputs, code interpreter, file search, MCP tools and background responses for both.OpenAI — OpenAI Chat Completions and Responses APIs.learn.microsoft.com/en-us/agent-framewo… ↗

Open models

YesOllama, Foundry Local and ONNX Runtime GenAI run models locally, with reduced capability: Ollama function tools depend on the chosen model and the provider table gives ONNX no tool support at all.Ollama — Local open-source model inference.learn.microsoft.com/en-us/agent-framewo… ↗

Model choice

YesYou choose the model: eleven providers are documented with a per-capability comparison table, and in .NET any client implementing the IChatClient abstraction can back an agent.Any inference service that provides a Microsoft.Extensions.AI.IChatClient implementation can back a ChatClientAgent.learn.microsoft.com/en-us/agent-framewo… ↗

Where you use it

In your editor

NoNot an editor product. The documented client surfaces are OpenAI-compatible HTTP, A2A, AG-UI web apps, Telegram and MCP; no IDE or editor extension appears in the docs.After choosing a host, select the client integration that fits your scenario:learn.microsoft.com/en-us/agent-framewo… ↗

On the command line

PartlyThe devui command launches agents discovered from a directory with flags for port, host, headless, tracing and auth, but DevUI is documented as a sample app 'not intended for production use'.If you have agents and workflows organized in a directory structure, launch DevUI from the command line:learn.microsoft.com/en-us/agent-framewo… ↗

In a browser

YesDevUI is a dev-time web interface with an OpenAI-compatible backend and trace viewer; for shipping your own web front end the documented integrations are AG-UI event streaming and ChatKit.It provides a web interface for interactive testing along with an OpenAI-compatible API backend, allowing you to visually debug, test, and iterate on…learn.microsoft.com/en-us/agent-framewo… ↗

Not established: In your pipeline

Whose machine it runs on

Self-hosted

YesSelf-hosting is one of the two documented hosting models; your application owns routing, identity, authorization, storage, deployment and scaling. The self-hosting packages are prerelease in both Python and .NET.Self-hosting lets you run an Agent Framework agent or workflow in your own web application, container, service, or runtime.learn.microsoft.com/en-us/agent-framewo… ↗

Open source

YesMIT-licensed on GitHub, with the Go SDK in a separate microsoft/agent-framework-go repository; the docs site itself is open to public contributors.Yes, Agent Framework is open source and available on GitHub.learn.microsoft.com/en-us/agent-framewo… ↗

What it costs to run

How it meters

PartlyMicrosoft charges nothing for the framework itself; every unit of cost is metered by whichever model provider or Azure service you point it at, and the vendor assigns that cost to you explicitly.You are responsible for any usage and associated costs.learn.microsoft.com/en-us/agent-framewo… ↗

Free tier

YesNo call, no account with Microsoft: the packages are MIT-licensed and installable from PyPI and NuGet. You still need credentials for a model provider, or a local runtime.Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to…github.com/microsoft/agent-framework/bl… ↗

API access

YesIt is a library you call directly, and the hosting helpers expose an agent or workflow over HTTP: OpenAI Responses and Chat Completions-compatible endpoints, A2A, or your own routes.OpenAI-compatible endpoints for Responses and Chat Completions-compatible APIs.learn.microsoft.com/en-us/agent-framewo… ↗

MCP server

PartlyExposing your agent or workflow as an MCP tool is Python-only today: .NET support is 'coming soon' and Go is 'not currently available'. Consuming MCP servers as tools works in all three languages.Use agent-framework-hosting-mcp to expose an Agent Framework agent or workflow as a tool on the native Model Context Protocol SDK.learn.microsoft.com/en-us/agent-framewo… ↗

Bring your own key

YesYou bring the inference: your provider account and credentials (Azure identity, provider API key) or a local runtime such as Ollama, Foundry Local or ONNX. Microsoft supplies no inference with the SDK.Model providers supply the inference client used by an Agent Framework agent.learn.microsoft.com/en-us/agent-framewo… ↗

Buying it for a team

Admin controls

PartlyOrganisation-level policy needs the Purview middleware plus Microsoft Purview and an M365 E5 licence; everything in the framework itself (tool approval, ShellPolicy allow/deny lists, FIDES flow policies) is set by the developer in code, not by an admin.Prevent sensitive data leaks: Inline blocking of sensitive content based on Data Loss Prevention (DLP) policies.learn.microsoft.com/en-us/agent-framewo… ↗

Audit log

PartlyAdministrative audit comes from the Purview integration, which requires Purview and an M365 E5 licence; the framework on its own emits OpenTelemetry traces, logs and metrics of agent and tool activity, which is telemetry rather than an admin trail.Enable governance: Log AI interactions in Purview for Audit, Communication Compliance, Insider Risk Management, eDiscovery, and Data Lifecycle…learn.microsoft.com/en-us/agent-framewo… ↗

What happens to your code

Opt out of training

PartlyThere is no Microsoft-side switch because the SDK runs in your process and sends data only where you point it; training use is governed by the model provider's terms, and the vendor tells you to review third-party retention and handling yourself.We recommend reviewing all data being shared with and received from Third-Party Systems and being cognizant of third-party practices for handling,…learn.microsoft.com/en-us/agent-framewo… ↗

Data residency

YesYou hold the data: it runs in your process or your host, so residency follows your infrastructure and chosen provider, and Microsoft assigns that boundary decision to you in writing.It is your responsibility to manage whether your data will flow outside of your organization's Azure compliance and geographic boundaries and any…learn.microsoft.com/en-us/agent-framewo… ↗

Getting out

YesConversation and session state live wherever you put them, in a custom history provider or session store over your own database, Redis or blob storage, so there is no vendor-held data to export and no credits that can expire.For database/Redis/blob-backed history, implement a custom history provider.learn.microsoft.com/en-us/agent-framewo… ↗

Not established: Certifications