agentscope
AgentScope is an open-source Python framework for building and running agents "you can see, understand and trust". Version 2.0 is designed around increasingly capable models: rather than constraining them with rigid prompts and orchestrations, it gives them good building blocks and lets their reasoning and tool use do the work. It is developed by the agentscope-ai team and backed by an arXiv paper.
It suits Python developers building production agents and multi-agent systems. Building blocks include a ReAct agent loop with structured output and interruption and resume, a toolkit spanning Python tools, MCP servers and skills (with built-in coding tools), models from major providers including OpenAI, Anthropic, Gemini, DashScope, DeepSeek and Ollama, automatic context compaction, and a unified event stream for front ends. Recent releases add pipelines and team pipelines in which a leader delegates to member agents, the A2A protocol, model routing, standard operating procedures and realtime voice agents.
Getting started: requires Python 3.11+. Install with `uv pip install agentscope`, or install from source in editable mode, then build an agent from the ReAct building block and add tools and models.
Features
Why choose it
Trade-offs
Maintenance check
- Last commit
- 2026-10-23
- Commits, last 90 days
- 100+
- Releases, last 90 days
- 6
- Latest release
- v2.0.10 (2026-10-12)
- Open issues
- 697
- Recent issues with at least one comment
- 9/20
Oldest week on the left; the latest four weeks are orange. Source: GitHub commit activity.
Status rule: archived on GitHub = Archived; last commit within 90 days = Active; 91 to 365 days = Slowing; older = Stale. All figures come from the GitHub API on the date shown.
Most discussed open issues
- #926 [Feature]: human in loop wanted17 comments · 2025-11-12
- #2707 [Bug]: streamed Base64 accumulation repeatedly transcodes the full payload16 comments · 2026-09-20
- #3194 [Bug]: Provider length-limit stop reasons are lost, so incomplete tool calls may run15 comments · 2026-10-09
- #1120 [human in loop]how to interrupt the agent flow and give a confirm/new information to the agent to continue the flow when coding by fastapi?15 comments · 2026-01-15
- #2566 feat(realtime): support context management for the realtime agent14 comments · 2026-09-09
Issue titles are quoted from GitHub. Open the link to check whether a problem still applies to the version you use.
Who it fits
- You want composable Python 3.11+ building blocks (ReAct loop, toolkit, pipelines, A2A, realtime voice) with bilingual documentation.
- You use Chinese model providers (DashScope, DeepSeek, Volcengine) alongside global APIs under an Apache-2.0 license.
- You rely on v1.x AgentScope code and do not want to migrate to the redesigned v2.0 architecture.
- You need Python 3.10 or .NET support; AgentScope 2.0 requires Python 3.11+.
How it differs
AgentScope (Apache-2.0) emphasises transparent ReAct loops, A2A, realtime voice, and first-class support for Chinese and global LLM APIs. CrewAI (MIT) emphasises role-based Crews paired with event-driven Flows.
AgentScope is a general-purpose, actively updated agent toolkit; MetaGPT models a fixed software-company SOP (PM, architect, engineer) and has had slower repository activity since January 2026.
Compatibility
Quick start
Use cases
Alternatives
Featured in these lists
Comments
No comments yet. Be the first!