Grok Bot: your agent needs a stop button

An agent that makes progress while you sleep is tempting. With Grok Bot, xAI offers agents that can use applications and follow routines. The interesting question comes just after the demo: what can it do without asking you?

Preparing a prospect list and contacting those people are different jobs. So are proposing a fix and publishing it. Access to a tool should never leave that distinction implicit.

Start with a task that ends

For a first trial, I would choose an easy-to-review result: a draft report from approved documents. No sending, no payments, no access to the entire inbox. A dedicated account also avoids passing along the permissions accumulated on your personal account.

The contract below describes a policy for a service you build. It is not a configuration file understood by Grok Bot. Limits must be enforced by the server and tool permissions, not just written in a prompt.

1{
2  "task": "draft-weekly-report",
3  "allowedActions": ["read-approved-documents", "save-draft"],
4  "requiresApproval": ["send-email", "publish", "delete"],
5  "maxSteps": 12,
6  "maxDurationSeconds": 600,
7  "maxSpendUsd": 2,
8  "stopAfterConsecutiveErrors": 2
9}

Plan the stop before the start

Check authorization before each action. For the budget, reserve the estimated maximum cost of the next call and reject it if there is not enough left. Adding spending afterwards helps with tracking but does not prevent overspending. With concurrent tasks, that reservation must be atomic.

A stop button must block further actions and revoke access when necessary. It cannot undo an email already sent. For sensitive operations, keep a unique identifier so that retrying after a failure does not trigger the same action twice.

Be able to understand what happened

The report should show actions, approvals and errors without copying secrets or all personal data into logs. If you cannot explain why an agent acted, giving it more autonomy will mostly make the problem harder to trace.