AI Agent

Claude Code Architecture: Four Core Designs of Modern AI Agents Originally published in Chinese on 2026-04-01; this English edition preserves the original scope and technical context. Claude Code, Anthropic’s AI programming agent released in February 2025 and internally known as Tengu, evolved rapidly from its first v0.2 beta release to v2.1.88. Over that period it introduced background agents, an automatic permission mode, MCP server integration, and multi-agent collaboration through Agent Teams. Unlike LangChain and Google ADK, which present themselves as frameworks, Claude Code is an end-user product. It does not primarily offer an SDK for building other agents; it acts as a programming partner in the terminal. That positioning creates a different set of architectural constraints: framework-level extensibility and composability matter less, while security, context management, streaming interaction, and error recovery must meet production standards. ...

LangChain and LangGraph Architecture: From Chains to Graph-Driven Agent Orchestration Originally published in Chinese on 2026-03-03; this English edition preserves the original scope and technical context. LangChain is one of the most widely used open-source frameworks in the LLM application development domain, while LangGraph is the second architectural attempt by the LangChain team in the agent orchestration layer—a stateful workflow engine inspired by Google Pregel and Apache Beam. Their relationship is not one of simple replacement but rather complementary at different layers: LangChain provides model abstractions, tool encapsulations, and high-level Agent interfaces, whereas LangGraph handles the execution orchestration, state persistence, and human-in-the-loop control at the lower layers. ...

Google ADK Architecture: Building AI Agent Systems with Software Engineering Principles Originally published in Chinese on 2026-02-27; this English edition preserves the original scope and technical context. ADK (Agent Development Kit) is an AI agent development framework that Google open-sourced in 2025. It has more than 18,000 GitHub stars and implementations for Python, TypeScript, Go, and Java. Its design goal is explicit: bring agent development back into the discipline of software engineering instead of leaving it at the prompt-engineering stage. That requires clear abstraction layers, composable modules, deterministic execution flows, and production-grade state management. ...