AutoGen AgentChat

AutoGen AgentChat is the high-level API of Microsoft’s AutoGen framework for agentic AI. It gives you preset Agents and Teams with predefined multi-agent design patterns so you can build multi-agent LLM applications without wiring the low-level event system yourself.

Built on top of the autogen-core event-driven runtime, AgentChat is the recommended starting point for most users, while advanced users can drop down to core for finer control over the underlying components.

What You Get

• Preset Agent types with configurable behaviors and tools
• Team abstractions implementing proven multi-agent design patterns
• Termination conditions to control when agent conversations end
• A high-level API layered cleanly over the autogen-core runtime
• UI and streaming helpers for observing agent interactions

Common Use Cases

• Building assistant + tool-executor agent pairs that collaborate on tasks
• Orchestrating teams of specialized agents with defined roles
• Prototyping agentic workflows before dropping to the core runtime
• Adding conversational multi-agent reasoning to LLM applications

Under The Hood

Architecture - AgentChat (python/packages/autogen-agentchat/src/autogen_agentchat/) is organized into agents, teams, conditions, base, state, tools, and ui subpackages. Agents encapsulate LLM behavior; teams coordinate multiple agents using design patterns; conditions decide termination. It depends directly on autogen-core, delegating the event-driven message routing and component runtime to that layer.

Tech Stack - Python (>=3.10), built with hatchling, part of a uv-managed monorepo. Its only hard dependency is a pinned autogen-core, keeping the high-level layer thin. Tooling includes ruff, pyright, pytest with coverage, and poe task runners shared across the monorepo.

Code Quality - The package is strictly typed (pyright with reportDeprecated), has a dedicated tests/ suite run under pytest with coverage, and follows the monorepo’s shared lint/type configuration. Clear subpackage boundaries separate agents, teams, and conditions.

API Design - The public API favors intuitive defaults: instantiate preset agents, group them into a team, set a termination condition, and run. This makes the common multi-agent cases a few lines while leaving autogen-core available for advanced event-driven control.