AgentScope 2.0
A Python framework for building, operating, and deploying observable, controllable agents.
- Source repo
- agentscope-ai/agentscope
- Stars
- ★ 36k
- Last updated
- 2d ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 70/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A Python team building a terminal coding assistant can equip an Agent with shell, search, file-reading, writing, and editing tools while retaining confirmation and interruption controls.
- Not a fit if
- Projects stuck on Python 3.10 or older that cannot upgrade their runtime
- Teams wanting a ready-made chatbot rather than assembling their own model and tool stack
- Source review
- 70/100 · Some gaps
What does this agent do, and when should you use it?
AgentScope 2.0 is a production-oriented Python agent framework whose SDK includes Agent, a ReAct loop, Toolkit, model adapters, context management, an event system, permissions, middleware, and long-term memory. Its unified event bus streams reasoning, tool calls, and text, image, or audio content to frontends, while the runtime supports interruption, resumption, and sequential or concurrent tool execution. The model layer covers OpenAI, Anthropic, Gemini, DashScope, DeepSeek, Moonshot, Volcengine, xAI, and Ollama, along with embedding and text-to-speech interfaces. Tool and code execution can be placed in local, Docker, Apple Container, Bubblewrap, E2B, OpenSandbox, Daytona, or Kubernetes workspaces. A FastAPI agent service and pre-built Web UI extend the SDK with multi-tenancy, isolated sessions, agent teams, RAG, persistence, resource sharing, scheduling, and background work. It is best suited to engineering teams that want to assemble their own model, tool, permission, and deployment stack rather than adopt a fixed chatbot application.
Developers construct an Agent from a system prompt, chat model, and Toolkit; the supplied example uses DashScopeChatModel, reads credentials from DASHSCOPE_API_KEY, and gives an agent named Friday access to Bash, Grep, Glob, Read, Write, and Edit. Calling launch_console(agent) opens a streamed terminal conversation with tool-call confirmation and Ctrl+C interruption handling. The ReAct layer runs the reasoning-and-acting loop, supports structured output and realtime pause/resume, and can batch tool actions sequentially or concurrently. Toolkit manages Python tools, MCP servers, and skills, while context middleware performs automatic compaction, tool-result offloading, and injection of system prompts, RAG, or memory. The event system emits reasoning, calls, and multimodal content, and the permission/HITL layer controls access to tools and resources. In deployment, the FastAPI service manages tenants, sessions, state, teams, retrieval, schedules, and channel messages; RealtimeAgent connects to DashScope, OpenAI, Gemini, and xAI realtime APIs, while A2AAgent communicates with remote A2A agents.
- A Python team building a terminal coding assistant can equip an Agent with shell, search, file-reading, writing, and editing tools while retaining confirmation and interruption controls.
- A platform team launching a multi-user agent product can use the FastAPI service, pre-built Web UI, session isolation, persistence, and background task support.
- An application developer with a fixed business workflow or several specialist roles can use Pipeline or a leader–worker Agent Team with tracked task planning.
- A product team building speech interaction can use the experimental RealtimeAgent with DashScope, OpenAI, Gemini, or xAI realtime APIs.
- A team grounding responses in organizational knowledge can combine context injection, long-term memory, and the multi-tenant RAG service with blob storage and indexing.
- An engineering team running higher-risk tools can select local, Docker, Bubblewrap, E2B, Daytona, OpenSandbox, or Kubernetes execution and apply permission checks.
How do you install or deploy this agent?
Python 3.11 or newer is required. Install the package from PyPI:
uv pip install agentscopeFor an editable source installation:
# Pull the source code from GitHub
git clone -b main https://github.com/agentscope-ai/agentscope.git
# Install the package in editable mode
cd agentscope
uv pip install -e .The first-run example below uses DashScope, so a valid API key must be available as DASHSCOPE_API_KEY. The supplied material does not show credential setup for the other supported model providers.
How do you use this agent?
Create an agent with local coding tools and launch its terminal console:
from agentscope.agent import Agent
from agentscope.console import launch_console
from agentscope.tool import Toolkit, Bash, Grep, Glob, Read, Write, Edit
from agentscope.credential import DashScopeCredential
from agentscope.model import DashScopeChatModel
import os, asyncio
async def main() -> None:
agent = Agent(
name="Friday",
system_prompt="You're a helpful assistant named Friday.",
model=DashScopeChatModel(
credential=DashScopeCredential(
api_key=os.environ["DASHSCOPE_API_KEY"]
),
model="qwen3.6-plus",
),
toolkit=Toolkit(
tools=[
Bash(),
Grep(),
Glob(),
Read(),
Write(),
Edit(),
]
),
)
# Chat with the agent in the terminal — streamed output, tool-call
# confirmation and Ctrl+C interruption are all handled for you
await launch_console(agent)
asyncio.run(main())To run the bundled service, start the FastAPI backend first:
git clone -b main https://github.com/agentscope-ai/agentscope.git
cd agentscope/examples/agent_service
# start the agent service backend
python main.pyStart the Web UI in a second terminal:
cd agentscope/examples/web_ui
# start the webui
pnpm install
pnpm devWhat are this agent's strengths and limitations?
- The repository provides a broad path from ReAct, tools, context, and permissions to a FastAPI service, Web UI, RAG, and scheduling.
- Explicit adapters for OpenAI, Anthropic, Gemini, DashScope, DeepSeek, Moonshot, Volcengine, xAI, and Ollama reduce dependence on one model vendor.
- A unified event bus streams reasoning, tool activity, and text, image, or audio content for observable user interfaces.
- Permission checks, confirmations, bypass mode, realtime interruption, and resumption provide several layers of execution control.
- Execution targets span local processes, containers, hosted sandboxes, and Kubernetes, allowing teams to choose an isolation boundary.
- Python 3.11 or newer is mandatory, so projects on older Python versions must upgrade their runtime.
- The full Web application requires separate Python backend and pnpm-based frontend processes, adding deployment and operational overhead.
- The quickstart depends on a DashScope API key; although many providers are listed, their credential setup and feature differences are not demonstrated.
- Realtime voice-agent support is explicitly experimental, creating a risk of interface or behavior changes.
- Multi-tenant RAG, persistence, channels, scheduling, and numerous sandbox choices create a substantial configuration surface, with no production capacity, latency, or recovery measurements supplied.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| AgentScope 2.0 This agent | 70 · Some gaps | Library / SDKFree + model costs | ★ 36k | 2d ago | Python | OpenAI API · Claude API |
| Neuron AI — Agentic Framework for PHP | 45 · Major gaps | Library / SDKFree + model costs | ★ 2.2k | today | PHP | OpenAI API · Claude API |
| AgenticX Multi-Agent Platform | 64 · Some gaps | CLIFree + model costs | ★ 997 | today | Python | Claude Code · OpenAI API · Claude API |
| Synapse AI — Multi-Agent Orchestration Platform | 70 · Some gaps | CLIFree + model costs | ★ 337 | 9d ago | Python | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows the framework provides a permission system (Permission & HITL) and sandbox isolation (Workspace/Sandbox), supporting tool-level permission control and user confirmation, but lacks default least-privilege configuration or mandatory confirmation, hence 2. Data flow transparency: event system streams reasoning and tool calls, but data flow and storage locations are not clearly documented, hence 2. Sensitive data handling: API keys are passed via environment variables, but no encryption or redaction is mentioned, hence 1. Dependency security: dependency list is clear but no vulnerability scanning or lock files, hence 1. External effects: tools can execute shell commands and file operations, but controlled via permission system, hence 2. Rollback: no mention of state rollback or undo mechanisms, hence 1. Source attribution: license and third-party code attribution are clear, hence 2.
Self-consistency: README and pyproject.toml descriptions are consistent, tests show internal consistency, hence 2. Dependency availability: dependency list is clear with optional extras, but no lock files, hence 2. Failure messages: tests show error handling (e.g., invalid timezone fallback), but no user-friendly error message documentation, hence 2.
Audience and scenarios: README clearly targets developers and researchers, with multiple scenarios (agent service, RAG, team collaboration), hence 3. Capability boundaries: documentation lists building blocks and features, but no explicit limitations or boundaries, hence 2. Trigger precision: event system provides precise event types, but trigger conditions are not documented, hence 2. Environment fit: supports multiple Python versions and OS, but no containerized deployment instructions, hence 2.
Information architecture: README is well-structured with rich documentation links, hence 3. Install notes: provides PyPI and source installation steps, hence 3. Naming stability: API naming is consistent, but no naming convention documentation, hence 2. Examples and FAQ: provides multiple examples and documentation links, but no FAQ, hence 3. Known limitations: no explicit list of known limitations, hence 1. License: Apache-2.0 license is complete, hence 3. Versioning and changelog: has NEWS.md but no detailed changelog, hence 2. Maintenance responsibility: has contribution guide and community channels, but no explicit maintainers, hence 2.
Output usability: provides event streams and structured output for easy integration, hence 3. Marginal value: provides rich building blocks and features, offering unique value compared to other frameworks, hence 3. Cost-benefit: open-source and free, but deployment and running costs are not documented, hence 2.
Claim traceability: feature claims in README have documentation and example links, hence 2. Cross-source corroboration: has arXiv papers and documentation, but not independently verified, hence 2. Fact-inference separation: documentation distinguishes feature descriptions and examples, but does not explicitly separate facts and inferences, hence 2.
- No default least-privilege configuration; users must configure the permission system themselves.
- Sensitive data handling (e.g., API keys) lacks encryption or redaction measures.
- Dependencies lack lock files, posing supply chain risks.
- Known limitations are not explicitly listed, potentially affecting production deployment decisions.
- No rollback mechanism provided; state changes may be irreversible.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's Changed * fix(model): declare real WAV sizes in the final streamed audio block by @BlueX888 in https://github.com/agentscope-ai/agentscope/pull/3242 * fix(embedding): isolate Ollama embedding caches by host by @maahumada in https://github.com/agentscope-ai/agentscope/pull/3265 * fix(model): await provider stream cleanup on public aclose by @HsbcJone in https://github.com/agentscope-ai/agentscope/pull/3216 * fix(workspace): offload nested media from context archives by @wuyak in https://github.com/agentscope-ai/agentscope/pull/3201 …View on GitHub ↗ -
Release notes excerpt
What's Changed * chore(version): bump version to 2.0.10dev by @DavdGao in https://github.com/agentscope-ai/agentscope/pull/2894 * docs: announce SOP, TeamPipeline and MiniMax support in news by @DavdGao in https://github.com/agentscope-ai/agentscope/pull/2932 * fix(tts): label audio with the effective response format by @little-spoon77 in https://github.com/agentscope-ai/agentscope/pull/2980 * fix(model): emit the dashscope audio transcript as text by @aniruddhaadak80 in https://github.com/agentscope-ai/agentscope/pull/2925 …View on GitHub ↗ -
Release notes excerpt
Highlight - Standard Operating Procedures (Experimental): Added the agentscope.sop module for splitting a task into ordered milestones, each verified before the next one starts, with run state that can be persisted and resumed. The agent service also exposes /sop endpoints to store procedures, start runs, and review steps by hand. - Team Pipeline: Added a TeamPipeline class where a leader agent delegates tasks to member agents, which run concurrently and hand their results back. …View on GitHub ↗