OpenAI Codex 云端编程 Agent 上线

本地终端里的编程 Agent 占一段连续时间：人提出任务，看它搜索与修改，遇到岔路就介入。屏幕亮着，注意力也被占着。2025 年 5 月 16 日，OpenAI 在 ChatGPT 里放出云端 Codex 研究预览，时间安排换了：任务被送走，各自进入隔离环境，几项工作可以同时跑，完成后再带着补丁和证据回来。人不必盯住每一步，却必须在返回时面对可检查的材料。

连接 GitHub 仓库后，Codex 为每个任务准备预装代码的容器，在其中读文件、跑命令、改代码、执行测试。首版由 codex-1 驱动。一个 Agent 看不到另一个任务在改什么；隔离让并行成为可能，也把任务边界写清楚。修 bug、补测试、解释一段代码，可以是三份独立委托，而不是在一个对话里排队。开放形态是研究预览，面向具备相应 ChatGPT 权限的用户，不是默认给所有人的“云工程师”。云端隔离降低对开发者笔记本的直接风险，却对所连接仓库与密钥注入方式提出新的治理问题。

云端没有把软件工程变成一句命令。Agent 仍可能误读需求，测试仍可能覆盖不足，仓库里未写下的约定仍可能被忽略。OpenAI 把日志、测试结果与代码引用放进交付物，让人能追查做过什么，而不是只接收完成宣言。任务在后台跑，责任没有被托管。十个同时失败的任务，只会更快制造十份噪声。并行数量不是能力本身。

这与终端 Agent 并非互斥：终端适合同步校正，云端适合边界清楚、可隔离、最后可验收的工单。它与 2021 年的 Codex API 模型同名不同物——那一代把自然语言变成可调用代码；这一代把软件任务变成可并行、可带回证据的云端作业。首版交付的不是“并行工程师”，而是一种更克制的产品单位：独立环境、过程记录与返回包。队列管理——哪些任务够独立、什么叫通过、冲突如何回主分支——开始和写代码本身一样占时间。

研究预览把“证据包”设为默认交付的一部分：不只是 diff，还有跑过什么、测过什么、引用了哪些位置。人不在场时，唯一能替代盯梢的就是可检查轨迹。与 Claude Code 等终端 Agent 对照，云端版用隔离与并行换取不同节奏；与 2021 年 Codex 对照，它不再满足于生成片段，而试图在仓库级任务上闭环。

A coding agent in a local terminal occupies a continuous stretch of time: a person assigns work, watches search and edits, and intervenes at forks. The screen stays lit; attention stays occupied. On 16 May 2025, OpenAI’s cloud Codex research preview inside ChatGPT rearranged that time. Tasks leave, each enters an isolated environment, several can run at once, and patches return later with evidence. People need not watch every step, yet must face checkable materials on return.

After a GitHub repository is connected, Codex prepares a container preloaded with code for each task, reads files, runs commands, edits, and executes tests. The first version was driven by codex-1. One agent cannot see what another is changing; isolation makes parallelism possible and writes task boundaries clearly. Fix a bug, add tests, explain a passage—three independent commissions rather than a single conversation queue. Access was a research preview for ChatGPT users with the right entitlements, not a default “cloud engineer” for everyone. Cloud isolation lowers direct laptop risk and raises governance questions about connected repos and injected secrets.

The cloud did not turn software engineering into one sentence. Agents still misread requirements; tests still under-cover; unwritten repository conventions can still be ignored. OpenAI put logs, test results, and code citations into the delivery so users can reconstruct what happened rather than receive only a completion claim. Tasks run in the background; responsibility is not hosted with them. Ten simultaneous failures only manufacture ten noise packets faster. Parallelism count is not capability.

The relationship to terminal agents is complementary: terminals suit correction-in-flight; cloud tasks suit work that is bounded, isolatable, and finally checkable. This product must not be confused with the 2021 Codex API model that shared the name—that generation turned natural language into callable code; this one turns software tasks into parallel cloud jobs that return evidence. The first preview did not deliver “parallel engineers.” It delivered a restrained product unit: own environment, process record, and return package. Queue management—which tasks are independent enough, what counts as pass, how conflicts re-enter main—begins to take as much time as writing code itself.

Research preview made an “evidence pack” part of default delivery: not only a diff, but what ran, what tested, which locations were cited. When people are absent, the only substitute for watching is a checkable trail. Set against terminal agents such as Claude Code, the cloud version trades isolation and parallelism for a different tempo; set against 2021 Codex, it is no longer content to emit fragments, but tries to close the loop on repository-scale tasks.