How to create a custom Grok agent, step by step.

Terms first

What custom Grok agents actually are.

In Grok Bot — xAI's agent product, in beta since 11 August 2026 under the tagline “AI teammates that finish the work” — an agent is a Bot: a persistent, named agent with its own cloud computer. It keeps working, and its routines keep running, while your machine is off. Custom Grok agents are simply Bots you configure yourself instead of copying from someone else's template.

Not to be confused with Grok companions (the avatar personas in the consumer app) or the older Grok Tasks feature in the chat app — more on those in the FAQ. The apps run on macOS, Windows, and iOS, and Grok Bot is bundled into all SuperGrok and Cursor plans, with a free limited trial.

Step 01

Name a Bot and write its instructions.

Create a new Bot and give it one job, in the name: “Ops,” “Inbox,” “Research.” Then write instructions the way you would brief a new teammate — the outcome you want, the constraints that matter, the tone to use, and when it should stop and ask you. A Bot holds its instructions, context, and constraints between sessions, so anything you write here compounds.

A useful test from our submission criteria applies here too: say what outcome the Bot creates, not that it is “powerful.” Vague instructions produce vague agents.

Step 02

Teach skills by demonstrating the work once.

Skills are reusable instruction sets a Bot applies on demand. You can write one by hand, but the faster route is “Teach a Task”: demonstrate the workflow once — the steps, the tools, the checks — and the Bot captures it as a skill it can repeat. Teach the awkward, specific things: how your team formats a changelog, which report goes to whom, what “done” means for a ticket.

Step 03

Put repeat work on scheduled routines.

Routines are scheduled or event-triggered workflows: a metrics summary every Monday morning, a triage pass when a GitHub issue lands, a reply drafted when a Slack thread mentions your product. Triggers cover time schedules plus Slack, GitHub, and Teams events. Because routines run on the Bot's own cloud computer, they fire whether or not your laptop is open.

Practical limits, as xAI has reported them: up to 50 routines per Bot, with the 20 most recent runs kept. Start with one or two scheduled routines you will actually read the output of, and add event triggers once the schedule earns its keep.

Step 04

Connect the accounts the job needs — and no more.

The first-party connectors named in xAI's docs and guides are Gmail, Google Drive, Google Calendar, Notion, Slack, Salesforce, HubSpot, and GitHub, with plugins and MCP servers beyond those. Bots request permissions conversationally as work requires them, which makes it easy to say yes — so decide up front which actions need a human approval: sending, publishing, deleting, purchasing, and production changes can all be gated.

A custom agent with narrow access that asks before consequential actions will outlive a broad one you stop trusting.

Step 05

Grow one Bot into an agent team.

When one Bot starts wearing too many hats, split the job. An agent team puts two to six Bots in a group chat, and the common pattern is a chief of staff: one Bot takes the goal, breaks it down, and delegates the pieces to specialists — a researcher, a builder, a reviewer. Keep each specialist's instructions narrow; the chief of staff is where the judgment lives.

The shortcut

Grok coding agents and other head starts.

You can also skip the blank page entirely: adding a public template creates an independent copy of someone else's working setup — instructions, skills, routines — on your account. It is the fastest way to see how experienced builders structure Grok coding agents, chief-of-staff Bots, and scheduled publishing pipelines. Four curated starting points from our directory: