agentscope

Implements a closed-loop reasoning-acting pattern where the LLM decides on tool calls, a Toolkit executes them, and results are integrated back into working memory for the next reasoning step. The architecture composes pluggable Model (OpenAI, Anthropic, Gemini, DashScope, Ollama), Formatter (provider-specific API payload conversion), Memory (working + optional long-term), and Toolkit components, enabling flexible agent behavior without strict prompt constraints.

Manages message broadcasting and coordination between multiple agents through a MsgHub component that automatically routes messages to enrolled participants. Supports predefined pipeline patterns (sequential_pipeline, fanout_pipeline) for complex multi-agent workflows where agents communicate asynchronously and decisions flow through the system. Built on top of the Msg abstraction, enabling agents to exchange structured messages with content blocks.

Supports model optimization through reinforcement learning (RL)-based fine-tuning and prompt tuning. RL fine-tuning allows agents to optimize their behavior based on reward signals, improving decision-making over time. Prompt tuning optimizes prompt templates without modifying model weights. Model selection capabilities enable choosing the best model for specific tasks based on performance metrics.

Provides realtime voice agent capabilities through integration with text-to-speech (TTS) models and audio streaming. Agents can process audio input, reason about it, and generate spoken responses in real-time. The architecture supports streaming audio for low-latency interactions and integrates with realtime model backends that support audio I/O natively.

Provides an evaluation framework for assessing agent performance across multiple dimensions (accuracy, efficiency, safety, user satisfaction). Evaluators can be custom-defined or use built-in metrics. The framework supports batch evaluation of agent trajectories, enabling systematic performance comparison across different agent configurations, models, or strategies.

Provides a PlanNotebook abstraction for structured task planning and decomposition. Agents can break down complex tasks into subtasks, track progress, and reason about dependencies. PlanNotebook integrates with the agent's memory and reasoning loop, enabling agents to maintain and update plans as they execute tasks.

Provides native integration for the Model Context Protocol, allowing agents to discover and invoke standardized external tools through HttpStatelessClient (for stateless tool calls) or StatefulClientBase (for tools requiring session state). The Toolkit component manages both local functions and MCP-based tools, exposing them to the ReActAgent through a unified interface. Formatters handle conversion of tool schemas into provider-specific function-calling formats.

Enables AgentScope agents to communicate with external agent systems across the network using the A2A protocol, allowing agents to discover, invoke, and coordinate with agents outside their local system. Agents can send messages to remote agents and receive responses, facilitating distributed multi-agent systems where agents may be built on different frameworks or deployed independently.

Provides a unified Msg object that encapsulates agent communication with support for multiple content block types (text, images, audio, structured data), enabling agents to exchange multi-modal information. The Msg abstraction decouples message structure from provider-specific formats, allowing Formatters to convert messages into OpenAI, Anthropic, Gemini, or other provider APIs. Content blocks are composable, allowing rich message payloads with mixed content types.

Converts unified Msg objects into provider-specific API payloads through a pluggable Formatter architecture. Each provider (OpenAI, Anthropic, Gemini, DashScope, Ollama) has a corresponding Formatter that handles protocol-specific details like function calling schemas, tool_use blocks, and message role mappings. Formatters are automatically selected based on the configured Model, enabling seamless provider switching.

Provides a dual-memory architecture where working memory holds the current conversation context (message history) and optional long-term memory persists information across sessions using configurable backends (vector stores, databases). Working memory is automatically managed by agents, while long-term memory requires explicit integration through memory retrieval/storage operations. Session management enables agents to resume from previous states or maintain persistent knowledge across multiple conversations.

Manages both local Python functions and remote tools (MCP, A2A) through a unified Toolkit interface. Functions are registered with schemas that enable the ReActAgent to discover and invoke them. The Toolkit handles function execution, error handling, and result formatting. Tool schemas are automatically converted by Formatters into provider-specific function-calling formats (OpenAI, Anthropic, etc.), enabling agents to decide when and how to call tools.

Provides RAG capabilities through a Knowledge Base abstraction that integrates with vector stores (Chroma, Pinecone, etc.) and document readers for ingesting and retrieving information. Agents can query the knowledge base to augment reasoning with relevant documents or facts. The architecture supports embedding models for semantic search and document chunking strategies for optimal retrieval.

Provides built-in observability through OpenTelemetry integration, enabling agents to emit traces, metrics, and logs for monitoring and debugging. Traces capture agent reasoning steps, tool calls, and model invocations, providing visibility into agent behavior. Metrics track performance (latency, token usage, tool call counts). Logs capture detailed execution information. OTel exporters send data to observability backends (Jaeger, Datadog, etc.).