AgentScope bets on smarter models, not smarter prompts

What it does

AgentScope 2.0 is a Python framework for building, running, and deploying LLM agents. It bundles ReAct agents, toolkits, memory, planning, realtime voice, evaluation, and model finetuning into one toolkit, and ships with a FastAPI-based service that supports multi-tenancy, multi-session chat, and a pre-built Web UI.

The interesting bit

Instead of forcing developers into prompt-engineering gymnastics, AgentScope treats the model as the orchestrator—leveraging native reasoning and tool-use capabilities. It also supports MCP and A2A protocols, so your agents can plug into broader tool and agent ecosystems without bespoke adapters.

Key highlights

• Streaming event architecture with EventType deltas for real-time agent feedback
• Deploys locally, serverless, or on Kubernetes with built-in OTel support
• MCP and A2A protocol support for tool and agent interoperability
• Includes model finetuning support and evaluation utilities out of the box
• Pre-built Web UI and multi-session FastAPI service for immediate prototyping

Caveats

• The README examples center on DashScope/Qwen models; broader model provider support is implied but not demonstrated in the provided source
• Requires Python 3.11+, which may sideline older environments

Verdict

Worth evaluating if you want a batteries-included path from prototype to production without writing a custom orchestration DSL. Skip it if you need a minimal agent library or are pinned to an older Python stack.