OpenAI has shipped persistent task execution in Codex via a new 'goal mode' command, activated with /goal, now out of experimental status and available across the Codex app, IDE Extension, and CLI.
The core mechanic: give Codex a milestone, and it runs until the work is done, spanning hours or days if needed. Users retain control through check-in points, steering corrections, and mid-run pauses. This is not a chatbot loop. It is an autonomous execution layer with human override built in.
The full video is worth watching for how goal mode handles long-horizon tasks without losing context or requiring constant prompting. The graduation from experiment to general availability signals OpenAI considers the reliability bar cleared. The question the video does not fully answer: what happens when Codex hits an ambiguous decision point mid-run with no user present.
[WATCH ON YOUTUBE →]