AutoGen

The Conversation Model#

AutoGen's central idea is that multi-step problem-solving can be expressed as a conversation between agents with different roles, tools, and prompts. Instead of a single agent invoking tools in a loop, you decompose the task across specialised agents — a planner, a coder, a critic, a tool user — and let them exchange messages until a termination condition is met.

This is a strong fit for tasks where decomposition into roles is natural: software engineering (planner + coder + reviewer), data analysis (analyst + executor + verifier), research (searcher + summariser + fact-checker). The original AutoGen paper (Wu et al., 2023) showed that multi-agent conversations often outperform single-agent baselines on coding, maths, and tool-use benchmarks.

AutoGen 0.4 Architecture#

The 0.4 rewrite, released in early 2025, was a significant break from the 0.2 line. It introduces a layered architecture and an async event-driven runtime suitable for long-running agents and distributed deployments.

• Core — the low-level message routing runtime. Agents are addressable actors; messages are typed events. Suitable for advanced users building custom topologies.
• AgentChat — the high-level API most users interact with. Provides `AssistantAgent`, `UserProxyAgent`, `CodeExecutorAgent`, and team patterns (`RoundRobinGroupChat`, `SelectorGroupChat`, `Swarm`).
• Extensions — integrations with model providers, code executors (Docker, local, Jupyter), and tools.
• AutoGen Studio — no-code visual designer that compiles to AgentChat configurations.

Built-in Team Patterns#

AgentChat provides three team patterns out of the box, which cover most production multi-agent designs:

Code Execution Agents#

One feature that distinguishes AutoGen from peers is first-class support for code-executing agents. A `CodeExecutorAgent` can be wired to a Docker-backed Python or shell executor, letting the agent write and run code in a sandbox as part of the conversation. Combined with a planner-coder-reviewer team, this is how AutoGen's reference implementations tackle Math-AI and coding benchmarks.

When to Pick AutoGen#

Pick AutoGen when your problem decomposes naturally into multiple agent roles with distinct prompts and tools, especially if one of those roles is a code executor. Pick CrewAI for a more opinionated role-based team API. Pick LangGraph if you want fine-grained graph-level control over a single complex agent. Pick OpenAI Agents SDK if you have already committed to OpenAI's ecosystem and want first-party support.

AutoGen Studio in particular is worth knowing about for organisations where non-engineers need to prototype or audit agent workflows — visual configuration with a typed underlying schema is rare in this space.