agentscope

What AgentScope 2.0 is

AgentScope 2.0 is a production ready, easy to use agent framework with essential abstractions that keep up with rising model capability. The README says the project is designed for increasingly agentic large language models. The approach leans on the models' own reasoning and tool use abilities rather than constraining them with strict prompts and opinionated orchestration. That framing positions AgentScope as a layer that gives structure without fighting the model.

The page lists a stream of recent feature and integration news from 2026. Pipeline support lets multiple agents run by fixed logic behind one event stream. Channels connect agents to instant messaging platforms, with DingTalk, Feishu, Lark, and Discord all added. A console mode lets developers test and debug agents in the terminal. MCP and Skill Hub support lets users browse a hub, install into a library, and add to a workspace, with the GitHub MCP Registry and ClawHub offered as built in hubs. Workspace and sandbox support spans Daytona, Kubernetes, OpenSandbox, and others, and ReMe long term memory is available. The news section is dated and points to example directories and documentation for each item, which signals an actively developed project rather than a static release. The 2.0 label itself indicates the framework has moved past its earlier 1.0 design toward these service oriented capabilities.

Building blocks of the SDK

The SDK layer composes an agent from a set of building blocks. ReAct provides a reasoning acting loop with structured output, realtime interruption and resume, and batched sequential or concurrent tool acting. Toolkit manages agentic tools over Python functions, MCP servers, and skills, and ships built in coding tools such as shell, file edit, and search plus task and plan tools. Model covers LLM, embedding, and text to speech across major providers including OpenAI, Anthropic, Gemini, DashScope, DeepSeek, Moonshot, xAI, and Ollama. Context handles automatic compaction, tool result offload, and context injection through built in middleware. An event system streams reasoning, tool calls, and multimodal content to the frontend over a unified bus. Permission and human in the loop control offers fine grained control over tools and resources, with confirmation and bypass modes. Middleware provides composable hooks across the loop. Memory is agentic with switchable backends such as ReMe and Mem0. Workspace and sandbox isolate tool and code execution across local, Docker, Apple Container, Bubblewrap, E2B, OpenSandbox, Daytona, and Kubernetes.

The README shows a first agent in the console: build an Agent with a name, system prompt, a model such as DashScopeChatModel, and a Toolkit of Bash, Grep, Glob, Read, Write, and Edit, then call launch_console to chat with streamed output, tool call confirmation, and interruption handled for you.

Agent service and deployment

Beyond the SDK, AgentScope ships a batteries included agent service: a FastAPI backend with a pre built web UI that turns agents into a multi tenant, multi session application. The serving capability gives multi tenancy, multi session isolation, a FastAPI backend, and the web UI. Agent Team provides leader worker orchestration with built in team tools and task planning. Channels connect agents to IM platforms such as Feishu, Discord, and custom channels with message routing. A RAG service offers blob storage, an index worker, and multi tenant retrieval. The MCP and Skill Hub lets users browse hubs like the GitHub MCP Registry and ClawHub, install into a library, and add to a workspace. Resource sharing manages models, MCP servers, skills, and workspaces at group and org level. Persistence stores agent state and sessions in SQL and NoSQL backends. Scheduling covers scheduled tasks, agent wakeup, and background task offloading.

Running it is straightforward. Clone the repository, run python main.py inside examples/agent_service to start the backend, then in a second terminal go to examples/web_ui, run pnpm install and pnpm dev to start the web UI. The README stresses that everything above is composable, so applications can be assembled with minimal glue code on top of the service.

Community, license, and research

AgentScope requires Python 3.11 or higher. Installation is shown from PyPI with uv pip install agentscope, and from source by cloning the main branch and installing in editable mode with uv pip install -e. The project welcomes community contributions and points to a CONTRIBUTING.md for guidelines. Community channels listed are Discord and DingTalk, with QR codes and invite links on the page.

The license is Apache License 2.0. The README also includes a publications section for researchers who find the work useful, citing two papers: AgentScope 1.0, a developer centric framework for building agentic applications, and AgentScope, a flexible yet reliable multi agent platform, both hosted on arXiv with full BibTeX entries. The first paper is listed with a 2025 Corr volume and the second with a 2024 volume, matching the arXiv identifiers shown in the badges. The combination of an Apache license, a documented SDK, a ready made service, and cited research positions AgentScope as both a practical tool and an academic project. The project also shows a contributors graph and links to its Chinese language README, so the audience is intentionally broad across regions and backgrounds. The steady stream of 2026 feature additions in the news section is a sign of active maintenance rather than a finished, static release.

Editorial conclusion

AgentScope is released under the Apache License 2.0 and requires Python 3.11 or newer, with source at github.com/agentscope-ai/agentscope.