源
Codex
cloud software engineering agent
BASEPowered by codex-1 — o3 fine-tuned via RL for software engineering; each task runs in an isolated sandbox preloaded with the user’s repo
AGENTS.md guidance
In-repo text file instructs the agent how to navigate code, run tests, and follow conventions
RL fine-tuning
Trained on real coding tasks — produces code closer to human style
Verifiability
terminal logs · test outputs provide an operational chain of evidence
Safety by design
Communicates uncertainty explicitly — users must review all generated code
Parallel tasks · sandbox isolation · auditable logs — the product-ready entry ticket for cloud SWE agents
→ codex · harness-engineering · openaiopenai.com/index/introducing-codex
Introducing Codex
- 来源:
sources/openai_official/introducing-codex.md - URL: https://openai.com/index/introducing-codex/
- 作者: OpenAI
概述
Codex 产品发布公告。云端软件工程 agent，支持并行多任务，基于 codex-1（o3 针对软件工程的 RL 微调版）。每个任务在独立沙箱中运行，预装用户仓库。
关键特征
- AGENTS.md 引导：repository 内的文本文件指导 agent 如何导航代码库、运行测试、遵循项目规范
- codex-1 模型：通过 RL 在真实编码任务上训练，产出更接近人类风格的代码
- 可验证性：通过 terminal logs、test outputs 提供操作证据链
- 安全设计：不确定时明确沟通，用户需审查所有 agent 生成的代码
与其他概念的关联
- Codex — 本文是产品发布公告
- Harness engineering — AGENTS.md 是 harness 知识的载体
References
sources/openai_official/introducing-codex.md