{"id":33560,"date":"2026-07-01T18:28:54","date_gmt":"2026-07-01T12:58:54","guid":{"rendered":"https:\/\/www.systweak.com\/blogs\/?p=33560"},"modified":"2026-07-02T11:20:17","modified_gmt":"2026-07-02T05:50:17","slug":"script-ai-agent-tasks-with-codex-gemini-claude","status":"publish","type":"post","link":"https:\/\/www.systweak.com\/blogs\/script-ai-agent-tasks-with-codex-gemini-claude\/","title":{"rendered":"How to Script AI Agent Tasks With Codex, Gemini (agy), and Claude"},"content":{"rendered":"
\n

If you automate work with CLI AI agents, you quickly hit the same wall: each agent takes its prompt, its working directory, and its role in a slightly different way, and the wrong shell or the wrong flag order silently breaks the whole run. This guide distills the rules that actually work on Windows, learned the hard way, so you can script a task once and hand it to whichever agent you prefer.<\/p>\n<\/div>\n

\n<\/span>Example user:<\/strong>\u00a0all paths below use a fictional Windows user\u00a0John Smith<\/strong> – short username jsmith<\/code>, long username\u00a0johnsmith<\/code> (no spaces). Swap in your own username where you see these.<\/div>\n
<\/div>\n<\/div>\n
\n


\nGolden rule we learned the hard way:<\/strong>\u00a0use\u00a0cmd<\/code>\u00a0(.cmd<\/code>\u00a0batch files) to launch these CLIs,\u00a0not<\/em>\u00a0PowerShell. The reason is stdin redirection and quoting (see Section 2).
\n<\/div>\n<\/span><\/p>\n<\/div>\n
\n

\n<\/span>\u26a0\u00a0Safety – the examples run agents unattended.<\/strong>\u00a0They use\u00a0--dangerously-bypass-approvals-and-sandbox<\/code>\u00a0(Codex),\u00a0--dangerously-skip-permissions<\/code>\u00a0(agy), and\u00a0--permission-mode bypassPermissions<\/code>\u00a0(Claude). These\u00a0turn off approval prompts and sandboxing<\/strong> so the agent can act on its own – Codex’s own help calls this “EXTREMELY DANGEROUS.” Use them only in a folder you trust, on work under version control that you can revert, never on a shared or production machine. Drop the flag (or use Codex -s read-only<\/code>\u00a0\/ agy\u00a0--sandbox<\/code>) when you want a safety net.
\n<\/div>\n<\/span><\/p>\n<\/div>\n

1. The Three Clients at a Glance<\/h2>\n
\n\n\n\n\n\n\n\n
Client<\/th>\nExe (portable form)<\/th>\nWorking dir set by<\/th>\nPrompt in<\/th>\nInstructions \/ params passed by<\/th>\n“Done” signal<\/th>\n<\/tr>\n<\/thead>\n
Codex<\/strong>\u00a0(OpenAI)<\/td>\n%LOCALAPPDATA%\\Programs\\OpenAI\\Codex\\bin\\codex.exe<\/code><\/td>\n-C \"<dir>\"<\/code>\u00a0flag<\/td>\n< prompt.txt<\/code>\u00a0(stdin)<\/td>\na\u00a0ROLE token<\/strong>\u00a0as one quoted positional arg<\/td>\nstate file ends\u00a0RUN_STATUS: COMPLETE<\/code><\/td>\n<\/tr>\n
agy<\/strong>\u00a0(Gemini)<\/td>\n%LOCALAPPDATA%\\agy\\bin\\agy.exe<\/code><\/td>\npushd \"<dir>\"<\/code>\u00a0cwd, or\u00a0--add-dir<\/code><\/td>\n-p \"text\"<\/code>\u00a0arg only – no stdin<\/strong>\u00a0(see Section 3, \u2020)<\/td>\nno token<\/strong>\u00a0\u2192 uses its default role<\/td>\nthe output file it was told to write exists<\/td>\n<\/tr>\n
Claude<\/strong>\u00a0(Claude Code)<\/td>\nclaude<\/code>\u00a0(must be on\u00a0PATH<\/strong>)<\/td>\npushd \"<dir>\"<\/code>\u00a0(uses cwd)<\/td>\n< prompt.txt<\/code>\u00a0(stdin)<\/td>\ncommand-line\u00a0flags<\/strong>\u00a0(--model<\/code>,\u00a0--effort<\/code>, \u2026)<\/td>\nstate file ends\u00a0RUN_STATUS: COMPLETE<\/code><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n