Skip to main content
Based on official documentation and community discussions (as of Feb 2026).

​
About AgentScope

AgentScope is a multi-agent framework designed to provide a simple yet efficient way to build LLM-empowered agent applications.
AgentScope v1.0 is a complete refactoring of the framework, equipped with new features and improvements. The most notable change is the shift to a code-first development model—the drag-and-drop Workstation UI from v0.x is no longer maintained or recommended for new projects.
Yes. You can refer to the Plan with ReAct Agent example on GitHub.
Absolutely. AgentScope’s clean codebase and tutorials are highly compatible with AI pair programmers. Provide the GitHub repo or tutorial source (docs/tutorial/en/src/) as context for best results.
Yes. Explore agentscope-samples for real-world use cases like Werewolf games, debates, and customer service bots built by the community.
  • AgentScope-Java is the Java port of AgentScope (under active development), aligned with the Python version in design and features.
  • Spring AI Alibaba will adopt AgentScope-Java as its underlying engine. If you use Spring AI Alibaba’s Agentic APIs, you’ll automatically gain AgentScope capabilities after upgrading—no need to integrate AgentScope-Java separately.

​
About Models

AgentScope has built-in support for DashScope, Gemini, OpenAI, Anthropic, and Ollama APIs, as well as OpenAIChatModel compatible with DeepSeek and vLLM models.
Create a custom model class by inheriting from agentscope.model.ChatModelBase and implement the __call__ method.
AgentScope Studio provides visualization of token usage and tracing. See the Observability section for details.
LLM providers (e.g., OpenAI, Anthropic, DashScope) have different input format requirements such as role names and tool call syntax. AgentScope uses Formatters to convert internal messages into API-compliant payloads, ensuring correctness even when vendors only partially support OpenAI-style APIs.
MultiAgentFormatter flattens multi-agent conversation history into a single string like "Alice: Hello\nBob: Hi" and sends it as a user-role message. This preserves global context but loses per-agent identity and tool semantics in the LLM’s view. It’s best suited for summary or coordination tasks—not fine-grained collaboration.
  • Fine-tuning updates model weights for better task performance.
  • Memory retrieval injects relevant context at inference time (e.g., via vector databases).
The two are complementary—you can even fine-tune a model to better leverage retrieved memories.

​
About Agents & Tools

You can use the ReActAgent class directly, or create a custom agent by inheriting from AgentBase or ReActAgentBase. See the Agent section for details.
  • agent_base: Abstract base class defining the core agent interface.
  • react_agent_base: Implements the ReAct (Reason + Act) paradigm, handling thought-action loops.
  • react_agent: A concrete, ready-to-use agent that extends react_agent_base with tool integration.
Yes. AgentScope supports dynamic JSON Schema generation and leverages Pydantic for validation in ReAct agents. See the Agent page for details.
Yes. Place your skill definitions in a directory following AgentScope’s skill structure, then register them via toolkit.register_agent_skill().
Use the pre-hook of the print function to forward printed messages. See the Hooking Functions section for details.
AgentScope includes a set of built-in tools such as execute_python_code, execute_shell_command, and write_text_file. You can find the full list under the agentscope.tool module.
Yes. AgentScope supports standard-compliant MCP for tool and service integration. See the Tool Capabilities page for details.

​
About AgentScope Runtime

Your current approach works well for single-agent, single-environment deployments. AgentScope Runtime is designed for more advanced scenarios:
  • Decouples business logic from execution environment via a unified protocol.
  • Enables independent upgrades of the runtime engine without touching agent code.
  • Supports elastic scaling of multiple agent instances (e.g., on Kubernetes).
If you only deploy fixed agents in one environment, your existing workflow is sufficient. Use Runtime when you need portability, scalability, or multi-platform deployment.
  • AgentApp defines the service interface (similar to FastAPI), using decorators such as @agent_app.query(framework="agentscope").
  • Runner implements the actual agent execution logic.
Requests sent to AgentApp are delegated to the Runner. The decorator-based pattern is the recommended approach—see the QuickStart guide.
The Sandbox provides an isolated, secure environment for executing code, file operations, or browser actions.
Never re-run sandbox-validated operations on the host system.
Instead, follow these safe practices:
  • Mount read-only host directories into the sandbox for data access.
  • Let agents write to a dedicated sandbox workspace.
  • Use sandbox outputs (e.g., downloaded files) directly in downstream steps.
The goal is zero risk to the host system.
Yes. Starting with v1.0.2, AgentScope Runtime will support CLI-based deployment, reducing boilerplate for Kubernetes and other platforms. The aim is to let you deploy the same agent artifact across ModelStudio, AgentRun, K8s, and more—without any code changes.

​
Reporting Bugs & Community

If you encounter a bug, open an issue on the AgentScope GitHub repository.
If you discover a security issue, report it through the Alibaba Security Response Center (ASRC).
Join the official community group through Discord or DingTalk:
DiscordDingTalk
DiscordDingTalk