Approvals and the overnight problem

The single most important safety setting on any personal agent is: what can it do without asking? Each of the three answers it differently, and you should know your agent's model before you let it run unattended.

• Meta Muse is the strictest: a separate Sentinel agent must approve anything before it reaches the internet, and Muse checks with you before sensitive actions like sending an email or buying, showing a full audit trail.
• OpenAI Dots uses a permission model plus Auto-review and reports back for review — though early users noted the occasional unapproved action, so keep write access tight until you trust it.
• Grok Bot uses Auto Review rules: Ask first always stops an action; Allow automatically lets it proceed unless review finds a reason not to; when both match, Ask-first wins.

Now the trap that catches everyone — the overnight problem. On Grok Bot, an approval for work that started without you (a routine, a trigger) expires after about 10 minutes. So a task that needs your tap at 3 a.m. doesn't wait until morning — it expires and does nothing. You wake up to a column of expired cards and no work done.

The fix is a mindset that works for all three: decide each unattended task's ending in advance. Every action it might take is either:

1. Safe and reversible → pre-authorize it (an Allow automatically rule on Grok Bot; granting the capability on Muse/Dots) so it finishes on its own.
2. Risky or irreversible (sending, publishing, buying, deleting, changing production) → design the task to stop at a draft or a recommendation and wait for you.

There is no good third option where the agent "waits indefinitely for me." Sort every action into allow-or-draft when you build the task, not at 3 a.m.