OpenAI Codex

Context

Reads workspace files, images, command output, AGENTS.md instructions, connected tools, and task history; cloud tasks start from a configured repository environment.

Actions

Edits code, runs commands and checks, performs reviews, calls configured tools, and can return a diff or pull request.

Execution

Local clients operate in configurable filesystem and network sandboxes; delegated tasks run in isolated cloud environments prepared from the repository setup.

Memory & rules

AGENTS.md supplies repository guidance. Threads can be resumed across connected surfaces; saved product context is separate from the underlying model weights.

Permissions

Sandbox modes limit filesystem and network access, while approval policies determine when commands may escape those boundaries. Organization policy can further restrict behavior.

Verification

Runs tests, linters, type checks, and Git inspection, then cites files and terminal output. A human still decides whether the evidence is sufficient and whether to merge.

Best for

• Switching between local pair work and delegated cloud tasks
• Repository changes that need a reviewable diff or PR
• Parallel investigation when tasks can be isolated

Limitations

• Editorial judgment: actions requiring credentials or services absent from the selected sandbox
• Editorial judgment: production changes without an explicit human release gate