AgentScope Builder

Problem to Solve

When building agents with AgentScope, a common failure mode is mixing APIs, directories, and prompts: the initialization order for ReActAgent, Toolkit, memory, and PlanNotebook is unclear; model and formatter are mismatched; tools lack docstring; skill documents contain descriptions but no usage rules. Such code may pass a demo but is hard to extend into multi-agent systems, persistent memory, or reusable skills.

How the Skill Works

The skill first classifies the request as a single agent, tool-driven agent, research agent, workflow agent, or skill-document need, then selects the smallest official abstraction set:

• Model layer: define model and formatter together; for OpenAI-compatible endpoints, use OpenAIChatModel with client_kwargs.
• Tool layer: register functions through Toolkit, and use create_tool_group() only when a capability domain needs explicit groups.
• Memory layer: start with InMemoryMemory(), then add persistent memory and memory-search tools when cross-session context is required.
• Planning and orchestration: use PlanNotebook for decomposable tasks and official pipeline for multi-agent flow instead of custom schedulers.
• Skill layer: encode repeatable rules in SKILL.md and attach them with register_agent_skill().

It also borrows CoPaw engineering patterns: a custom agent class wrapping ReActAgent, prompt files such as AGENTS.md, SOUL.md, and PROFILE.md, and runtime hooks for context checks or compression.

Boundaries and Notes

It fits engineering tasks that need runnable skeletons, directory layouts, and extension paths. It is not for conceptual overviews or frameworks outside the AgentScope API surface. The default output favors Python 3.10+ and async style; synchronous scripts, non-Python stacks, or arbitrary DAG scheduling require additional adaptation.