agentscope.go

AgentScope.Go is a production-oriented AI agent development framework implemented in Go, currently at version v2.6.0. It adopts the ReAct (Reasoning + Acting) paradigm, providing developers with all the necessary infrastructure for building intelligent applications based on large language models. Core features include: 1. Multi-model support: The framework has built-in adapters for mainstream model providers such as OpenAI (Chat Completions and Responses API), Anthropic Claude, Google Gemini, Alibaba Cloud DashScope (Tongyi Qianwen), DeepSeek, Moonshot (Kimi), xAI (Grok), vLLM, and Ollama. Any service compatible with the OpenAI API format can be integrated via BaseURL configuration. 2. Memory management system: Provides five memory implementations—InMemory, Window, ReMeInMemory, ReMeFile, and ReMeVector—supporting 7 vector backends, hybrid search (BM25 + Reranker), Dream evolution, and knowledge graphs. ReMe long-term memory supports both file-based and vector-based modes, automatically extracting personal/task memories and writing them to the vector store, with unified retrieval. 3. Tool system: Includes built-in file operations (Read/Write/Edit/Glob/Grep), Shell, Web, JSON, multimodal tools, as well as advanced tools like Task, Schedule, and Subagent. Supports custom FunctionTool with parameters defined via JSON Schema. 4. Multi-agent orchestration: Offers various orchestration patterns including Pipeline (sequential execution), Parallel (concurrency), Condition (conditional branching), Loop (iterative optimization), and MapReduce (long-document summarization), along with MsgHub broadcast-style message scheduling. 5. Control Plane: Supports long-horizon goal management, quota control, user gating, evidence verification, lease management, and dashboard views, with SQL persistence support, enabling a governance-evolution closed loop. 6. Multi-platform integration (Channel): Supports chatbot integration for Webhook, Discord, and Feishu platforms, providing unified message routing and reply mechanisms. 7. RAG hosted knowledge base: A complete document processing pipeline supporting Text/PDF/PPTX/Image/Word/Excel parsing, chunking, vectorization, indexing, and retrieval, with an HTTP API interface. 8. Terminal TUI (Console): A bubbletea-based three-state machine terminal interface, supporting HITL (human-in-the-loop) confirmation, Ctrl+C interruption, and three-level event rendering. 9. Hub marketplace: Supports browsing and installing MCP/Skill cards, with an FSHub implementation and zip-slip security protection. 10. Observability: Integrates OpenTelemetry, LangSmith, and Langfuse, supporting agent lifecycle tracking and event stream forwarding. 11. ONNX local inference: Pure Go implementation of image (CLIP) and audio (Whisper) preprocessing pipelines, connecting to ONNX Runtime services via HTTP proxy, with zero CGO dependencies. 12. A2A protocol: A complete Agent-to-Agent protocol implementation supporting AgentCard, Task, SSE, Registry, ShardRouter, as well as authentication, rate limiting, and WebSocket real-time communication. 13. Security and permissions: Multi-tenant session isolation, RBAC role permissions, rule engine, Bash compound command splitting, and HITL confirmation modes. 14. Production service Bootstrap: gateway.AppConfig + NewApp provides one-click production service assembly, supporting automatic tool injection, tool unloading, permission mode configuration, and embedding cache. 15. Plugin system: Three-stage lifecycle, YAML configuration, and Linux .so dynamic loading. 16. Message bus: LocalBus, RedisBus, and CoordBus with four primitives (Lock/Registry/Queue/Log), plus TeamBus for cross-session projection. 17. Self-evolution (Evolver): GEP Gene/Capsule types, Run/Reflect/Solidify workflow, and Skill-to-Gene distillation. The framework provides rich example code covering scenarios from basic agent usage to production-grade full-featured services, including A2A protocol, RAG Q&A, multi-agent orchestration, voice dialogue, multi-tenant workspaces, control plane, and ONNX inference. In terms of performance, embedding cache hits are approximately 550ns/op, cross-modal similarity is about 741ns/op, vector store search (1000 nodes) is about 229μs/op, full-text search (1000 documents) is about 97μs/op, and the gateway/ReAct hot path handles about 147k req/s. The environment requires Go 1.25 or higher.