AutoGen
Open source Microsoft framework for building multi agent AI apps with chat tool use function calling human in the loop and orchestration primitives for production workflows.
Overview
AutoGen provides a programmable way to compose multiple AI agents that can converse plan call tools and hand off work. Developers define agents with roles and capabilities then register functions APIs or code execution as tools the agents may request. A conversation loop coordinates messages and tool calls including explicit handoffs to a human when needed.
The framework includes examples for retrieval augmented generation data processing and coding assistants and supports both deterministic scripted exchanges and dynamic decision making. A companion set of extensions helps with memory logging evaluation and UI. Because AutoGen is open source teams can self host choose models and integrate with existing stacks while keeping observability through traces.
Microsoft now recommends its newer Agent Framework for many scenarios yet AutoGen remains maintained for fixes and remains widely used in research and prototypes that evolve into production when governance and controls are added.
Key features
- Agent Roles: Define planner executor critic or custom roles
- Tool Calling: Register Python functions APIs or shell tasks
- Conversation Loop: Coordinate agent messages tool calls and human handoffs
- Memory and Logs: Persist conversations and tool results for debugging
- Deterministic Scripts: Encode repeatable dialogues for reliability
- Extensible Storage: Plug in vector stores and retrieval sources
- Evaluation Helpers: Score runs with basic metrics and comparisons
- UI Starters: Use quick chat front ends to demo agents
Best for
- Customer Support Flows: Triage issues call CRM tools summarize tickets
- Data Processing: Pull files clean columns analyze and report
- Developer Copilots: Draft tests refactors open PRs with approval gates
- Research Assistants: Combine retrieval and critique roles with citations
- Operations Runbooks: Encode dialogues that escalate to humans with logs
- Marketing Drafts: Connect CMS analytics to propose briefs and drafts
- Back Office Tasks: Automate repetitive forms via internal APIs with logs
- Experimentation: Prototype multi agent ideas quickly then port patterns into newer frameworks
Capabilities
Roles and Policies
Declare clear agent responsibilities handoff points and fallback paths to keep behavior predictable and auditable.
Tool Calls
Expose safe functions so agents act in your environment while you validate inputs outputs and side effects.
Memory and Traces
Store interactions tool results and costs to debug failures and measure improvements over time.
Human in the Loop
Insert review and approval steps for risky actions so prototypes graduate to compliant production flows.
Frequently Asked Questions
Is it free to use?
Yes AutoGen is open source and free to use under its repository license on GitHub.
Does Microsoft still recommend it?
It is maintained with fixes while Microsoft points many new projects to the newer Agent Framework for greenfield builds.
Which models can I use?
You can call popular hosted or local models through adapters and standard Python clients.
Can agents run code?
You can register code execution and external tools with safeguards and resource limits.
Does it work with retrieval?
Yes extensions and examples show how to ground answers with vector stores and document loaders.
How do I monitor agents?
Use logging and tracing hooks to capture steps tools costs and outcomes for review.
Is there a UI?
Starter chat UIs and notebooks help teams demo flows before building custom front ends.
Can I deploy to production?
Teams often start in notebooks then containerize services with observability and policy controls for production.
