AutoGen

What it does

Writes code

YesA built-in CodeExecutorAgent generates code from an instruction and executes it, and Magentic-One ships a Coder agent specialised for writing code.An agent that generates and executes code snippets based on user instructions.microsoft.github.io/autogen/stable/refe… ↗

Debugs and fixes

PartlyWithin a run the agent retries failed executions and reflects on the result; nothing diagnoses a failure in an existing repository or lands a fix there.The maximum number of retries on error. If the code execution fails, the agent will retry up to this number of times.microsoft.github.io/autogen/stable/refe… ↗

Reviews code

PartlyThe reflection design pattern pairs a coder agent with a reviewer agent that critiques the generated code; no surface reviews a diff or pull request someone else wrote.For example, given a task to write code, the first LLM can generate a code snippet, and the second LLM can generate a critique of the code snippet.microsoft.github.io/autogen/stable/user… ↗

Whole-codebase work

PartlyFileSurfer navigates directories and reads local files while the executors run commands in a work_dir, so file-by-file work is possible; no repository indexing or repo-wide migration is documented.The FileSurfer can also perform common navigation tasks such as listing the contents of directories and navigating a folder structure.microsoft.github.io/autogen/stable/user… ↗

Ships it

PartlyAgents get a real shell through ComputerTerminal and Docker or local executors, so a deploy command can be run, but no deployment or release capability is documented for user code.ComputerTerminal provides the team with access to a console shell where the Coder’s programs can be executed, and where new programming libraries can…microsoft.github.io/autogen/stable/user… ↗

Done-for-you service

NoSoftware you install and run yourself; the project offers no implementation or managed-service engagement and support is community-managed through GitHub Discussions and Discord.AutoGen is a framework for creating multi-agent AI applications that can act autonomously or work alongside humans.github.com/microsoft/autogen ↗

Not established: Writes tests

How much it does unattended

Runs autonomously

YesTeams run to a termination condition with no human in the loop; the Magentic-One Orchestrator plans, delegates subtasks and revises the plan until the task is complete.the ability to autonomously adapt to, and act in, dynamic and ever-changing web and file-system environmentsmicrosoft.github.io/autogen/stable/user… ↗

Multi-agent

YesThis is the product's core: RoundRobinGroupChat, SelectorGroupChat, MagenticOneGroupChat and Swarm team presets, plus GraphFlow for directed-graph workflows.A team is a group of agents that work together to achieve a common goal.microsoft.github.io/autogen/stable/user… ↗

Agent permissions

PartlyPermission is code you write: an approval callback invoked before each code execution, a UserProxyAgent that blocks for human input, max_turns and termination conditions. Nothing is gated by default.A function that is called before each code execution to get approval.microsoft.github.io/autogen/stable/refe… ↗

Which models it runs on

Claude

Yesautogen-ext ships AnthropicChatCompletionClient over the Anthropic Python SDK; the docs head that section "Anthropic (experimental)".To use the AnthropicChatCompletionClient, you need to install the anthropic extra. Underneath, it uses the anthropic python sdk to access the models.microsoft.github.io/autogen/stable/user… ↗

GPT

YesOpenAIChatCompletionClient and AzureOpenAIChatCompletionClient cover GPT models, and every quickstart in the docs is written against one of them.To access OpenAI models, install the openai extension, which allows you to use the OpenAIChatCompletionClient.microsoft.github.io/autogen/stable/user… ↗

Open models

YesAn Ollama client runs open-weight models locally, and the Azure AI Foundry client is documented against Phi-4; the Ollama section is labelled experimental.Ollama is a local model server that can run models locally on your machine.microsoft.github.io/autogen/stable/user… ↗

Model choice

YesYou choose the provider: a model-client protocol in autogen-core with shipped clients for OpenAI, Azure OpenAI, Azure AI Foundry, Anthropic, Ollama, Gemini and Llama API.Since there are many different providers with different APIs, autogen-core implements a protocol for model clients and autogen-ext implements a set…microsoft.github.io/autogen/stable/user… ↗

Where you use it

In your editor

NoThe documented surfaces are the Python and .NET libraries plus two developer tools, Studio and Bench; no IDE or editor extension is offered.AutoGen Studio provides a no-code GUI for building multi-agent applications.github.com/microsoft/autogen ↗

On the command line

PartlyA built-in Console UI streams a run to the terminal and UserProxyAgent can take console input, but there is no standalone agent CLI; the autogenstudio command only launches the web UI.The Console() method provides a convenient way to print messages to the console with proper formatting.microsoft.github.io/autogen/stable/user… ↗

In a browser

YesAutoGen Studio is a browser UI with Team Builder, Playground and Gallery views, which you host yourself; the vendor states it is a prototyping tool, not production-ready.An web-based UI for prototyping with agents without writing code.microsoft.github.io/autogen/stable ↗

Not established: In your pipeline

Whose machine it runs on

Self-hosted

YesSelf-hosting is the only option: pip-installed packages, a standalone or distributed runtime across your own machines, and Studio backed by your SQLite or PostgreSQL database.Distributed runtime is suitable for multi-process applications where agents may be implemented in different programming languages and running on…microsoft.github.io/autogen/stable/user… ↗

Open source

YesCode in the public microsoft/autogen repository is MIT-licensed; the documentation content carries a separate Creative Commons Attribution 4.0 licence.grant you a license to any code in the repository under the MIT License, see the LICENSE-CODE filegithub.com/microsoft/autogen ↗

What it costs to run

How it meters

PartlyThe framework charges nothing; the only meter is your model provider's per-token billing, because you supply the account and the key.The following samples call OpenAI API, so you first need to create an account and export your key as export OPENAI_API_KEY="sk-...".github.com/microsoft/autogen ↗

Free tier

YesInstall and run it with no account, no card and no sales call; the only cost is your own model provider's API usage.Permission is hereby granted, free of charge, to any person obtaining a copy of this softwaregithub.com/microsoft/autogen/blob/main/… ↗

API access

YesThe product is itself an API: autogen-core, autogen-agentchat and autogen-ext on PyPI, Microsoft.AutoGen packages on NuGet, and Studio teams runnable from Python via TeamManager.AgentChat is a high-level API for building multi-agent applications.microsoft.github.io/autogen/stable/user… ↗

MCP server

PartlyAgents consume MCP servers over stdio, SSE and streamable HTTP through McpWorkbench, but the mcp module documents client sessions and tool adapters only, so an AutoGen agent is not itself queryable as an MCP server.McpWorkbench for using Model-Context Protocol (MCP) servers.microsoft.github.io/autogen/stable ↗

Bring your own key

YesEvery key and endpoint is yours: OpenAI, Azure OpenAI with key or Entra token auth, Anthropic, Azure AI Foundry, or a local Ollama server, all billed to you.You will also need to obtain an API key from OpenAI.microsoft.github.io/autogen/stable/user… ↗

Buying it for a team

Admin controls

PartlyConstraints are developer-level: the runtime enforces identity and privacy boundaries, code executes in a Docker sandbox, and an approval callback can gate each execution. There is no admin console or policy engine.the framework provides a runtime environment, which facilitates communication between agents, manages their identities and lifecycles, and enforce…microsoft.github.io/autogen/stable/user… ↗

Audit log

PartlyRuns are observable through OpenTelemetry spans for the runtime, tools and agents, plus Python event logging of model calls, but nothing records administrative action, because there are no administrators.AutoGen has native support for open telemetry. This allows you to collect telemetry data from your application and send it to a telemetry backend of…microsoft.github.io/autogen/stable/user… ↗

Single sign-on

NoThere is no authentication of any kind: Studio ships without login and the vendor tells developers to implement authentication themselves. SAML and OIDC appear nowhere in the docs.Developers are encouraged to use the AutoGen framework to build their own applications, implementing authentication, security and other features…microsoft.github.io/autogen/stable/user… ↗

What happens to your code

Getting out

YesTeams, agents and conditions serialise to JSON via dump_component and save_state, and Studio persists sessions in your own SQLite or PostgreSQL database. No credits exist to expire.State is a dictionary that can be serialized to a file or written to a database.microsoft.github.io/autogen/stable/user… ↗

Not established: Certifications