AutoGen Starter

Implements BaseGroupChat abstraction enabling multiple agents to communicate in structured conversation flows with configurable termination conditions and message routing. Uses AgentRuntime protocol to manage agent lifecycle, message subscriptions, and event propagation across agent instances. Supports round-robin, speaker selection, and custom routing strategies for coordinating agent interactions without explicit message passing code.

AssistantAgent wraps ChatCompletionClient to enable agents to call external tools via schema-based function registry with native bindings for OpenAI, Anthropic, and Ollama function-calling APIs. Integrates with CodeExecutorAgent for executing generated code in sandboxed environments. Agents maintain conversation history and can reason about tool outputs to refine responses iteratively.

Integrates with Model Context Protocol servers to discover and use tools via standardized MCP interface. Agents can connect to MCP servers (local or remote) and automatically discover available tools without hardcoding tool schemas. Tool calls are routed through MCP protocol, enabling interoperability with any MCP-compatible service. Supports resource access patterns for files, databases, and APIs.

GrpcWorkerAgentRuntime enables agents to execute on remote worker processes/machines via gRPC protocol. Central coordinator dispatches agent tasks to workers, collects results, and manages message routing across distributed agents. Supports horizontal scaling by adding more worker processes. Agents are location-transparent — same code runs locally or distributed without modification.

Enables capturing and persisting complete conversation state (messages, agent decisions, tool calls, results) to external storage for later analysis, debugging, or replay. Agents emit structured events that can be logged to databases, files, or observability platforms. Supports replaying conversations to reproduce issues or analyze agent behavior deterministically.

Enables agents to read files, write outputs, and interact with web resources (HTTP requests, web scraping) through sandboxed interfaces. Agents can fetch web content, parse HTML/JSON, and save results without direct file system access. Supports resource access patterns with permission controls and rate limiting. Integrations in autogen-ext provide implementations for common web/file operations.

Integrates memory systems (vector stores, knowledge bases) with agents via autogen-ext, enabling agents to retrieve relevant context before generating responses. Supports RAG patterns where agents query external knowledge sources, incorporate retrieved documents into prompts, and refine answers based on retrieved context. Memory systems are pluggable and can use different backends (in-memory, vector databases, custom implementations).

Implements agents that can learn from user feedback and examples during conversations, updating their behavior without retraining. Uses message history and feedback signals to refine agent responses iteratively. Agents can store learned patterns in memory systems and apply them to future interactions. Enables human-in-the-loop learning where agents improve through interaction.

Enables agents to pause execution and request human approval before taking actions (tool calls, code execution, final responses). Implements bidirectional communication where humans can override agent decisions, provide corrections, or inject new instructions mid-conversation. Uses message subscription and event routing to integrate human feedback into agent decision loops without blocking other agents.

CodeExecutorAgent provides safe code execution in isolated environments (Docker containers, local Python processes, or remote sandboxes) with resource limits and output capture. Agents can generate code, submit it for execution, and receive results without direct access to the host system. Supports multiple language runtimes and execution strategies configured via autogen-ext implementations.

ChatCompletionClient protocol abstracts LLM interactions across OpenAI, Anthropic, Azure OpenAI, Ollama, and custom providers. Agents use the same code regardless of underlying LLM provider. Supports configurable routing, fallback chains (try OpenAI, fall back to Anthropic if rate-limited), and provider-specific parameter mapping. Implemented via autogen-ext with native bindings for each provider's API.

Enables defining specialized agents with distinct roles (code reviewer, data analyst, researcher, executor) that collaborate on complex tasks. Each agent has configurable system prompts, tools, and capabilities. Teams are composed by instantiating multiple agents and connecting them via BaseGroupChat with role-aware routing. Agents can delegate tasks to teammates based on expertise.

Implements pluggable termination conditions that evaluate after each agent turn to determine if a group chat should end. Supports conditions like 'max turns reached', 'agent votes to stop', 'specific keyword detected', or custom logic. Termination is evaluated asynchronously without blocking other agents. Provides termination reason and final state to enable post-conversation analysis.

GraphFlow enables defining agent workflows as directed acyclic graphs (DAGs) where nodes are agents or tasks and edges represent dependencies. Agents execute in topological order with automatic parallelization of independent branches. Supports conditional branching based on agent outputs and dynamic task injection. Integrates with AgentRuntime for distributed execution across multiple workers.