What Goal mode is for

A new agent mode for giving AI a finish line instead of another prompt.

In May 2026, OpenAI released Goal mode in Codex, and Anthropic shipped /goal in Claude Code.

At first, you might read that and think the best use case is:

Build me my new website. Make no mistakes.

Then the agent disappears into another dimension and comes back with something over-designed, super well animated, and very proud of itself.

Well, you’d be wrong. That’s exactly the wrong use case.

Goal mode is for when the completion criteria is binary.

Normal prompting is:

Do this next.

Goal mode is:

Keep working until this condition is true.

The agent should be able to answer done or not done without asking for your taste, your opinion, or your patience.

That makes the best use cases… boring.

Is the error still thrown?
Does the project compile?

For the past 2 months, I used it once to fix an automation where the email notification kept failing. The run would report that the email was blocked, end the task, and I would discover the problem much later, which defeated the purpose.

I used goal mode to make the agent keep iterating on the integration until it could actually send the email.

Other good examples:

Process emails until the inbox is empty.
Categorize every expense receipt until none are left uncategorized.
Clean up calendar conflicts until there are zero overlapping meetings.

These tasks can technically work in normal mode. But when there are many items, models tend to stop a bit early and report accurate, but incomplete, progress.

Goal mode fixes that.