AgentScope
A production-grade intelligent agent development framework open-sourced by Alibaba. It supports agent construction, deployment, management, and execution, providing core capabilities including ReAct mode, tool auditing, human-agent collaboration, and sandbox isolation. Both Python and Java versions are available.
Core Facts
SSE or WebSocket protocols can push the model's token-by-token generated content to the front end in real time, enabling streaming responses from the Agent
Agent planning capabilities are typically implemented through the ReAct (Reasoning + Acting) framework.
AgentScope is an Agent development framework launched by Alibaba, with message exchange at its core and a built-in distributed execution mechanism.
Timeline (last 90 days)
AgentScope provides a variety of context management methods, including summarization and compression of historical messages, evicting old content by importance or recency, offloading large results to external storage and keeping only references, and truncating and structurally extracting tool returns, all made into composable modules.
The Bilibili uploader "智博AI" published an AgentScope 2.0 tutorial that it calls "the most complete and detailed on all of Bilibili," covering basic concepts and environment configuration through to two runnable conversation Agent examples.
The only difference between the second streaming Agent in the tutorial and the first instance is that stream is set to True when passing in the model, and Event is used to iterate over event blocks to check whether they contain a data element in order to decide what to print.
The first basic conversation Agent in the tutorial does not have streaming output enabled, and when the input "你好" is demonstrated, the model returns relatively slowly.
Running AgentScope requires the use of asynchronous programming (async/await): conversation methods must be defined with async def, await is used where the model is called, and at startup an asynchronous entry point such as asyncio.run() must be used; it cannot be called directly in a synchronous manner.
In AgentScope, when instantiating OpenAIChatModel, you pass in the API Key and the request address (base_url), can specify a concrete model name such as Qwen 3.6 Plus, and set the stream parameter to True to enable streaming output.
The role of AgentScope's Message has three values: User, System, and Agent, corresponding to the user, system, and assistant role conventions in the OpenAI Chat Completions protocol.
In AgentScope, Message is composed of Blocks: TextBlock corresponds to text and ImageBlock corresponds to images; the content field of Message is a list of Blocks, supporting multimodal messages.
In AgentScope, all interactions (between users and Agents, between Agents, and between Agents and RAG systems) are completed through Message.
AgentScope is a multi-agent development framework open-sourced by Alibaba's Tongyi Lab, and version 2.0 was released in May 2026.
40 more timeline events
All Facts (20)
SSE or WebSocket protocols can push the model's token-by-token generated content to the front end in real time, enabling streaming responses from the Agent
90%VerifiedAgent planning capabilities are typically implemented through the ReAct (Reasoning + Acting) framework.
90%VerifiedAgentScope is an Agent development framework launched by Alibaba, with message exchange at its core and a built-in distributed execution mechanism.
80%VerifiedAgentScope's three-dimensional integrated security defense line consists of three layers—tool review, human-machine collaboration, and sandbox isolation—addressing the questions of what can be done, what humans permit, and where it can be done safely.
75%VerifiedAgentScope provides tool invocation, streaming execution, and interrupt-and-resume capabilities centered around ReAct.
75%VerifiedWhen authorizing an Agent to perform high-risk operations, a manual confirmation step should be established, where the Agent proposes the plan and humans retain final execution authority (human-in-the-loop mechanism).
70%VerifiedAgentScope is a multi-agent framework developed and open-sourced by Alibaba's Tongyi Lab, used for building multi-agent applications in complex scenarios.
70%VerifiedAgentScope is an agent development framework aimed at Java developers, providing a toolchain from core reasoning to engineering deployment.
65%VerifiedAgentScope offers both a Python version and a Java version.
65%VerifiedStreaming execution allows the Agent to output intermediate results in real time, and checkpoint & resume ensures that long-running tasks can continue from the last interruption point when a failure occurs
65%VerifiedAgentScope 2.0 has reached production-level usage standards.
65%VerifiedAgentScope provides multiple context management strategies, including sliding window truncation (retaining only the most recent N rounds of conversation).
65%UnverifiedFor developers just getting started, it is recommended to learn AgentScope 2.0 directly rather than starting from 1.0.
60%UnverifiedAgentScope provides a variety of context management methods, including summarization and compression of historical messages, evicting old content by importance or recency, offloading large results to external storage and keeping only references, and truncating and structurally extracting tool returns, all made into composable modules.
50%UnverifiedIn AgentScope, Message is composed of Blocks: TextBlock corresponds to text and ImageBlock corresponds to images; the content field of Message is a list of Blocks, supporting multimodal messages.
50%UnverifiedThe role of AgentScope's Message has three values: User, System, and Agent, corresponding to the user, system, and assistant role conventions in the OpenAI Chat Completions protocol.
50%UnverifiedIn AgentScope, when instantiating OpenAIChatModel, you pass in the API Key and the request address (base_url), can specify a concrete model name such as Qwen 3.6 Plus, and set the stream parameter to True to enable streaming output.
50%UnverifiedAgentScope is a multi-agent development framework open-sourced by Alibaba's Tongyi Lab, and version 2.0 was released in May 2026.
50%UnverifiedIn AgentScope, all interactions (between users and Agents, between Agents, and between Agents and RAG systems) are completed through Message.
50%UnverifiedRunning AgentScope requires the use of asynchronous programming (async/await): conversation methods must be defined with async def, await is used where the model is called, and at startup an asynchronous entry point such as asyncio.run() must be used; it cannot be called directly in a synchronous manner.
50%