Untitled

This document focuses on routing, orchestration and multi-agent collaboration. It explains core design and practices, details the “Multiple Lobsters, One LinkMind” topology, provides hands-on guidance, and helps enterprises build unified, efficient, secure and compliant multi-agent systems.

This document covers agent safety design: threat models, bounded reachability, identity/roles, prompt injection defense, and the LinkMind security gateway. A lab hardens the assistant using role matrices, red-teaming, content filtering, capability gating, and audit. Also covers incident response and change management.

This document covers injecting private capabilities via skills, tools, and MCP, distinguishing read/write operations and side effects, emphasizing capability contracts (idempotency, observation shaping, identity/permissions). A lab builds an asset support layer (lookup owner, create ticket) and covers human approval, reversible workflows, versioning, and quota governance.

This document explains injecting private knowledge without retraining, focusing on the RAG pipeline (chunking, metadata, vector/hybrid retrieval, reranking) and grounded generation with refusal behavior. An internal policy assistant lab demonstrates corpus building, freshness testing, and provenance, with evaluation metrics and anti-patterns.

本文档介绍了如何构建可靠的单代理，内容涵盖最小代理定义、思维-行动-观察循环、隐藏职责（工具/模型治理、可追溯性）以及中间件的必要性。文中通过一个旅行办公桌实验演示了配置、工具调用、观察塑造和故障处理，并提供了评估标准、反模式和设计原则。