AGENTS.md serves as a project specification for AI intelligences, similar to an onboarding manual for human developers. When this file exists in the project root directory, Codex can quickly adapt to specific technical environments and development specifications, improving execution efficiency by more than 40%. This Markdown file utilizes a hierarchical directive system that supports multiple levels of granularity, from global specifications to module-level details.
The content of the file consists of three key modules: test specification requirements (e.g., pytest execution parameters), code style constraints (following PEP8 or Google Style), and a description of the project architecture (boundaries of responsibility for key modules). The deeply nested AGENTS.md takes precedence over directives, e.g., formatting requirements in subdirectories override common specifications in the root directory. As an example, a financial project configured with 5 layers of AGENTS.md increased the probability of Codex-generated code passing a code audit from 62% to 89% in a single pass.
Of particular note, when AGENTS.md contains automated validation scripts, Codex forces these checks to be performed before committing. An IoT project successfully blocked 3 code commits with possible injection vulnerabilities by integrating a customized security_linter.sh.
This answer comes from the articleChatGPT Codex: Intelligence for Automating Software Engineering TasksThe
English 
简体中文 
日本語 
Deutsch 
Português do Brasil