📄 中文摘要
OpenAI 最近的文档强调了代理工作流的概念，标志着从代码补全向完整代理运行时的转变。这种转变不仅涉及工具、逻辑节点、追踪评分和数据集，还引入了代理 SDK，形成了一种与传统“从提示生成函数”模型截然不同的抽象方式。随着技术的发展，编程的重心从简单的令牌预测转向更复杂的编排。编程的难点不再仅仅是局部语法，而是状态管理、工具使用、评估和故障后的恢复。一旦代理能够调用工具、检查输出、根据条件分支并将追踪反馈到评估循环中，软件问题的性质将发生显著变化。
将编码代理视为运行时系统
出处: Coding agents as runtime systems
发布: 2026年3月25日
📄 English Summary
Coding agents as runtime systems
OpenAI's recent documentation highlights the shift from code completion to full agent runtimes, emphasizing agent workflows that involve tools, logic nodes, trace grading, datasets, and the Agents SDK. This represents a significant departure from the traditional 'generate a function from a prompt' model, moving the focus from token prediction to orchestration. The complexities of serious coding are rarely about local syntax; instead, they involve state management, tool usage, evaluation, and recovery from failures. Once an agent can call tools, inspect outputs, branch on conditions, and feed traces back into evaluation loops, the nature of software problems begins to change dramatically.