Microsoft Agent Framework

What it does

Writes code

PartlyAgents you build can write and run code through CodeAct or a provider-hosted code interpreter, but both are gated: the documented CodeAct connector is in preview and code interpreter depends on the model provider you choose.CodeAct lets an agent solve a task by writing code and executing it through an `execute_code` tool. Instead of asking the model to emit one tool call…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

NoMicrosoft ships a package you install and run yourself; nobody builds or operates the agent for you, and support is the public GitHub repository, 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. It…learn.microsoft.com/en-us/agent-framewo… ↗

Not established: Debugs and fixes, Writes tests, Reviews code

How much it does unattended

Runs autonomously

YesThe agent-mode provider ships an explicitly autonomous execute mode that works a plan to completion, and the Harness keeps an agent progressing through long multi-step tasks without a person each turn.Execute is autonomous. The agent works through the plan, makes reasonable choices when details are ambiguous, and marks todos complete.learn.microsoft.com/en-us/agent-framewo… ↗

Multi-agent

YesFive built-in orchestration patterns coordinate several agents on one task, and a parent agent can additionally delegate concurrent work to named child agents in their own sessions.Agent Framework provides several built-in multi-agent orchestration patterns:learn.microsoft.com/en-us/agent-framewo… ↗

Agent permissions

YesApproval is per tool and configurable in code: the local shell tool asks for every command by default, regex allow and deny lists pre-filter commands, and disabling approval requires an explicit unsafe acknowledgement.`LocalShellTool` runs commands directly on the host. It defaults to a persistent shell, a 30-second timeout, 64-KiB output truncation,…learn.microsoft.com/en-us/agent-framewo… ↗

Which models it runs on

Claude

YesAnthropic is a first-class model provider with its own client package, reachable through Anthropic directly or through Foundry, Bedrock and Vertex AI endpoints.Anthropic — Claude model inference through Anthropic and supported hosted endpoints.learn.microsoft.com/en-us/agent-framewo… ↗

GPT

YesBoth OpenAI directly and Azure OpenAI are listed model providers, covering Chat Completions and the Responses API, with Microsoft Foundry as a third route to the same models.OpenAI — OpenAI Chat Completions and Responses APIs.learn.microsoft.com/en-us/agent-framewo… ↗

Open models

YesOllama, ONNX Runtime GenAI and Foundry Local are all listed as providers for running open-weight models locally, though the comparison table shows they support fewer hosted-tool features than the cloud providers.Ollama — Local open-source model inference.learn.microsoft.com/en-us/agent-framewo… ↗

Model choice

YesYou choose the model: eleven providers are documented and, in .NET, any inference service implementing the standard chat-client interface can back an agent without framework changes.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

PartlyA terminal surface exists only as developer tooling: the DevUI command line launches agents from a directory, and the Harness console apps are explicitly samples rather than shipped components.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

PartlyDevUI gives a browser interface for running and debugging agents, but the vendor labels it a sample app unfit for production; shipping a real web surface means adding the preview AG-UI or ChatKit integrations.DevUI is a **sample app** to help you visualize and debug your agents and workflows during development. It is **not** intended for production use.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: the agent runs in your own web service, container or runtime and you keep routing, identity, storage and scaling.Your application runs in your web service, container, runtime, or existing infrastructure.learn.microsoft.com/en-us/agent-framewo… ↗

Open source

YesThe whole framework is published under the MIT licence on GitHub, permitting commercial use, modification and redistribution without a fee or a contributor agreement.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… ↗

What it costs to run

How it meters

PartlyThe framework itself costs nothing, so all spend is model and tool consumption metered by whichever provider you configure; Microsoft publishes no calculator and states the usage cost is yours.Third-Party Systems are Non-Microsoft Products under the Microsoft Product Terms and are governed by their own third-party license terms. You are…learn.microsoft.com/en-us/agent-framewo… ↗

Free tier

YesNo sales call, no account and no licence fee: the packages install from PyPI and NuGet under an MIT grant that is explicitly free of charge.Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software")github.com/microsoft/agent-framework/bl… ↗

API access

YesThe product is itself a code library, and its hosting packages additionally expose an agent over OpenAI-compatible Responses and Chat Completions endpoints from your own server.OpenAI-compatible endpoints for Responses and Chat Completions-compatible APIs.learn.microsoft.com/en-us/agent-framewo… ↗

MCP server

PartlyAnother agent can reach yours over MCP or A2A, but the Python hosting packages that expose those endpoints are prerelease, and you must build and secure the server yourself.MCP tools for exposing an agent or workflow as a native MCP tool.learn.microsoft.com/en-us/agent-framewo… ↗

Bring your own key

YesYou bring your own inference: the framework requires you to configure your chosen provider's keys and endpoints, and it never brokers or resells model access.API Dependencies: Requires proper configuration of LLM provider keys and endpointsgithub.com/microsoft/agent-framework/bl… ↗

Buying it for a team

Admin controls

PartlyConstraints are real but developer-set in code, not an admin console: per-tool approval, regex shell allow and deny lists, and FIDES policy middleware that blocks a sensitive tool before it runs.`ShellPolicy` applies regular-expression allow and deny lists before execution. Deny rules take precedence.learn.microsoft.com/en-us/agent-framewo… ↗

Audit log

PartlyAgent actions are traceable via OpenTelemetry, and the FIDES security provider keeps per-session audit records of policy decisions, but there is no administrative action log because there is no admin plane.`SecureAgentConfig` stores labels, hidden variables, audit records, and pending approvals in the active `AgentSession`.learn.microsoft.com/en-us/agent-framewo… ↗

What happens to your code

Opt out of training

PartlyMicrosoft makes no training claim for the framework and instead pushes the question to whichever model or tool provider you configure, telling you to review their retention and privacy terms first.Provider Due Diligence: Before integrating any external service, you should review their privacy policies, security practices, data retention…github.com/microsoft/agent-framework/bl… ↗

Data residency

PartlyBecause you host it, residency follows your deployment and your chosen providers; Microsoft explicitly makes keeping data inside your compliance and geographic boundaries your responsibility rather than guaranteeing a region.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

YesNothing is locked in a vendor store: sessions and history go to a session store you implement over your own database or blob storage, and there are no credits to expire.Your application controls routing, identity, authorization, request policy, storage, deployment, and scaling.learn.microsoft.com/en-us/agent-framewo… ↗

Not established: Certifications