Seven Mental · 心智七篇
← Knowledge Atlas · Entity

Codex

OpenAI 的 agent 编码平台,隐式循环架构实现
ENTITY · CODEX · OpenAI · agent coding platform · harness-engineering at scale

Codex

OpenAI’s agent coding platform — powered by codex-1 (RL-finetuned o3), supports 6+ hour autonomous runs

Codex is OpenAI’s agentic coding solution, released in 2025 as a cloud agent product. Internally it runs an implicit-loop architecture (the Runner component drives prompt → tool call → reasoning → loop end). OpenAI teams used Codex to build a million-line product with zero hand-written code, validating harness engineering as the core of scaled agent development.

Core capabilities
Modelcodex-1: an RL-finetuned variant of o3, optimized for coding tasks
Long-runningSingle runs last 6+ hours, isolated sandbox per run, with parallel multi-task support
GuidanceAGENTS.md file guides agent behavior — analogous to Claude Code’s CLAUDE.md
Workflow integrationAutonomous PR opening, review response, and merge — full git-workflow automation
Harness engineering at scale
Zero-human million-line codebase
OpenAI team practice: a full production codebase requires no hand-written code
Uses standard tooling
gh, local scripts, in-repo skills — no proprietary API dependencies, validating tool-design principles
vs Claude Agent SDK
Both are agentic coding platforms running implicit-loop architectures — different emphases in their paths
→ Harness Engineering · Implicit Loop Architecture · Claude Agent SDKOpenAI Codex (2025)

Codex

OpenAI 的 agent 编码平台,支持 CLI 和云端运行模式。

概述

Codex 是 OpenAI 的 agentic 编码解决方案,类似于 Anthropic 的 Claude Agent SDK。其内部运行 隐式循环架构(Runner 组件驱动 prompt → tool call → reasoning → loop end 循环)。

Codex 于 2025 年作为 cloud agent 产品发布,基于 codex-1(o3 的 RL 微调版),支持并行多任务、独立沙箱、AGENTS.md 引导。

在 Harness Engineering 中的角色

OpenAI 团队用 Codex 零人工编码构建了百万行代码的产品(详见 harness engineering 实践),验证了 harness engineering 在规模化 agent 开发中的核心地位。关键特征:

  • 单次运行可持续 6+ 小时
  • 使用标准开发工具(gh、本地脚本、repo 内 skills)
  • 支持自主 PR 开启、review 响应、merge

相关实体

References

  • sources/openai_official/harness-engineering.md
  • sources/openai_official/unlocking-codex-harness.md
  • sources/openai_official/unrolling-codex-agent-loop.md