AgentScope
Type: full-code · Vendor: Alibaba ModelScope · Language: Python · License: Apache-2.0 · Status: active · Status in practice: emerging
Provide a production-ready agent framework with built-in ReAct agent, MsgHub for multi-agent message routing, MCP/A2A integration, real-time steering via hooks, structured output, and short/long-term memory.
Description. AgentScope is Alibaba ModelScope's Apache-2.0 Python framework billed as 'Build and run agents you can see, understand and trust.' It ships a ReActAgent out of the box (sync/async tools, streaming, hooks, structured output, configurable long-term memory), a Toolkit registry, MCP client transports (stdio / SSE / streamable HTTP), an MsgHub for multi-agent message routing, and OpenTelemetry-based observability for local, serverless, or Kubernetes deployment.
Agent loop shape. ReAct loop wrapped with hook points. A ReActAgent assembles formatted prompt + memory + registered tool schemas, the model emits reasoning plus tool calls or a final reply, tools are dispatched through a Toolkit (in-process functions or MCP-backed via stdio/HTTP/SSE clients), and results are appended to memory. Hooks fire pre/post each step, enabling interrupts and structured-output validation. Multi-agent topologies are expressed by chaining agents through an MsgHub that routes messages.
Primary use cases
- ReAct-style single-agent assistants with tool use
- multi-agent orchestration via MsgHub message routing
- MCP-integrated agents over stdio / SSE / streamable HTTP
- human-in-the-loop steering through real-time user interrupts
- OTel-instrumented production deployment
Key concepts
- ReActAgent → react (docs) — Out-of-the-box ReAct agent supporting sync/async tools, streaming responses, hooks, structured output, and configurable long-term memory.
- Toolkit → tool-use (docs) — Registry that registers tool functions for an agent; tools can be backed by MCP clients.
- MCP clients → mcp (docs) — HttpStatefulClient / StdIOStatefulClient / HttpStatelessClient transports; stateless clients create per-call sessions.
- MsgHub → orchestrator-workers — Message hub for flexible multi-agent orchestration and workflows.
- Long-term memory modes → cross-session-memory — agent_control / static_control / both — determine whether the agent self-manages memory or runtime auto-records.
- OTel observability — Built-in OpenTelemetry instrumentation for local, serverless, or Kubernetes deployment.
Patterns this full-code implements —
- ★★ReAct
ReActAgent is shipped as a top-level class; users register tools on a Toolkit and pass it to the agent.
- ★★Tool Use
Tools are registered on a Toolkit instance via register_tool_function(); supports sync and async tools.
- ★★Model Context Protocol
Stateful and stateless MCP clients for HTTP / streamable HTTP / SSE / stdio transports; register an entire server or individual functions.
- ★★Human-in-the-Loop
Hooks expose pre/post step points and the ReActAgent supports real-time steering with user interrupts.
- ★★Orchestrator-Workers
MsgHub routes messages between multiple agents for flexible multi-agent orchestration.
- ★★Structured Output
ReActAgent supports structured output as a documented capability of the agent.
- ★★Cross-Session Memory
long_term_memory_mode = agent_control | static_control | both — the agent persists memory across sessions either by self-management or runtime auto-record.
- ★Browser Agent
AgentScope ships a Browser-use Agent example linked from the README; not a first-class core abstraction.
- ★★Agentic RAG
RAG and Long-Term Memory are shipped as functionality examples linked from the README; no first-class RAG-Agent class documented in the quickstart.
- ★★Supervisor
The Meta Planner Agent example documents a planner that decomposes complex tasks and orchestrates sub-agents — AgentScope's closest analogue to a built-in supervisor.
- ★★Plan-and-Execute
Planning is listed among built-in capabilities; a Plan example and a Meta Planner Agent example ship from the README.
- ★Conversational Multi-Agent
Multi-agent message routing via MsgHub structurally resembles conversational multi-agent systems; AgentScope does not adopt the AutoGen ConversableAgent vocabulary. No verbatim upstream quote uses th…
- ·Debate
Multi-agent debate is shipped as a workflow example simulating a multi-turn discussion between several solvers and an aggregator.
- ★Agent-as-Tool Embedding
Agent-as-Tool is a documented multi-agent abstraction: agents can be registered as tools callable by other agents via the Toolkit class.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Listed as alternative by (1)
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: needs-verification