智能体执行框架的演进：从链到图再到代码
原标题：Article: The Evolution of Agentic Harnesses: From Chains to Graph to Code
AI 摘要
InfoQ 发表文章，梳理 agentic harness（智能体执行框架）从 Chains、Graphs 到 Code 三种范式的演进。文章指出，Chains 是线性工作流，构建快但脆弱，上游单点失败会级联；Graphs 通过节点和边引入显式控制流，支持条件分支、并行、状态管理和多智能体编排；Code harness 把执行当作一等原语，借助沙箱和迭代精炼实现确定性、可验证的输出。作者建议开发者将 LLM 输出视为不可信输入、解耦规划与执行、重视可观测性，并选择满足需求的最简抽象。
正文节选
Key Takeaways - The progression of agentic harnesses has unfolded across three distinct paradigms: Chains, Graphs, and Code, marking a transition from ad hoc, "best-effort" experiments to resilient, production-grade software architectures. - Chains represent the initial era of linear workflows, offering the lowest latency and compute costs for narrow tasks such as summarization, though they remain inherently fragile: a single upstream failure can cascade through the entire execution flow. - Grap