Grok Bot multi-agent setup

A Grok Bot multi-agent setup is not “make twenty custom bots.” It is a small desk: named seats, shared gates, and a weekly cadence. This walkthrough is for people already running a business with Grok Bot — not ChatGPT, not Claude, not a hosted agent platform. You paste markdown into Grok Bot custom / project instructions. The bots draft. You still click send, publish, and pay.

Ownrshp’s pack is Grok Bot Operator Fleet (OF-001) — $29. Follow the setup with your own prompts, or paste the zip.

What “multi-agent” should mean on a one-person desk

On a solo operator desk, agents are seats, not employees and not personalities you collect. Each seat owns one job and refuses the others:

• Chief of Staff — priorities, briefs, handoffs. Does not do Scout’s deep work.
• Market Scout — demand signals, angle cards, a short watch list. Does not publish.
• Product — SKU cards, pack outlines, honest listing copy. Does not mark live.
• Store Ops — listing fields, fulfillment notes, delivery QA. Never goes live.
• Research / Sales-draft — optional. Deep digs or long copy. Still gated. Never send.
• QA — a lane, not a seventh personality. Pass/fail against the rulebook before a human yes.

Four core seats, one optional specialist, a QA lane. Not twenty. Why lean: operator fleet vs random bots.

Templates optimize for more output. A multi-agent desk optimizes for ownership of the irreversible clicks.

Before you create the second bot

Write the shared rule down once. Every seat in a Grok Bot multi-agent setup must pin the same default deny. If only CoS has the gates, Product will cheerfully draft a “just publish it” and Ops will sound like it already did.

The three gates — explained as product philosophy in SEND / PUBLISH / PAY and as an agent-implementation checklist in agent SEND / PUBLISH / PAY gates:

• SEND — no emails, DMs, community posts, or support replies to real recipients.
• PUBLISH — no live listings, pages, directory submits, or go-live toggles. Staging is allowed.
• PAY — no charges, ads spend, refunds executed, crypto / wire / ACH.

Allowed without ceremony: drafts, tables, checklists, internal notes you paste back, recommendations. A recommendation is not execution. If you are unsure, treat it as gated. Draft + ask.

The 15-minute install order

Do this in order. Skipping gates “until later” is how random bots get born.

1. Open Grok Bot. Use custom / project / bot instructions — whatever your workspace calls the field you paste into. You are not creating a hosted login on ownrshp.io.
2. Create four bots with short names you will type all week: COS, SCOUT, PROD, OPS. Optional fifth: R&S.
3. Paste one seat per bot. Full instructions in that bot only. Do not dump every seat into one mega-prompt. Collision starts there.
4. Pin the always-on gates block in every seat (or as a shared project rule). Same wording. Not a paraphrase per bot.
5. Save three routines as named prompts: morning brief (→ COS), demand scan (→ SCOUT), weekly SKU pipeline (→ COS, then PROD).
6. Open a tracker — Notion, Sheets, or a markdown table — with a money/SKU row and a demand watch list. Agents should write into artifacts you keep, not into chat graveyards.

If you bought OF-001, those paste-ready files are 01–07 and 09 in the zip. If you are rolling your own, write the gates first, then CoS, then Scout. Product and Ops without gates is how a listing goes live in a thread you were using as a scratch pad.

First run: 90 minutes, not a prompt-decorating weekend

A multi-agent setup that you never operate is a folder. Run this once the same day you paste:

• 0–15 min — COS. Morning brief. Exit: top 3, one kill, one handoff. No essays. Copy the top 3 into your notes.
• 15–35 min — SCOUT. Demand scan on a niche you actually sell. Exit: three angle cards + a watch-list table (≤5 rows). Pick one angle. Defer the other two.
• 35–65 min — PROD. SKU card from the chosen angle. Exit: name, SKU, price, differentiator, file list, listing title / subtitle / five bullets. Honesty pass: no fake testimonials, no invented revenue.
• 65–85 min — OPS. Stage listing fields + delivery QA. Do not publish. Gate: PUBLISH.
• 85–90 min — you. Named yes or explicit stop. Full name, gate, artifact. Then you click in the real tool — or you do not, and it stays draft.

Done looks like five artifacts, not five new bots. Repeat with routines. Do not reward the first awkward handoff by inventing a sixth personality.

Handoffs that survive a Tuesday

Chat blobs are how work dies between seats. A handoff from CoS is one ask, one owner, and a “done looks like”:

To: SCOUT
Ask: Demand scan for [niche]. Three angle cards, no invented market sizes.
Done looks like: markdown table + one recommended angle for PROD.

CoS may recommend a publish. Ops may stage the listing. Neither marks live until a named human yes exists in the thread. Scout, Product, and Research/Sales never request payment tools. If an operator types “just publish it” without a name, every bot in the fleet should block, ask for the form, and refuse to invent a go-live.

Failure modes (and the fix)

• One mega-bot with seven jobs. Fix: split seats. CoS that secretly becomes Product will skip Scout and ship undifferentiated copy.
• Gates in CoS only. Fix: pin the same block everywhere. The blast radius is the bot you forgot.
• Twenty templates, no cadence. Fix: three routines, one SKU in draft per week unless you say otherwise. Prefer finishing.
• Agents that claim they already sent it. Fix: the bot must never narrate a completed SEND / PUBLISH / PAY. Draft + gate label + you click.
• Building this on ChatGPT or Claude and expecting the same desk. Fix: this setup is Grok Bot. Ownrshp does not maintain a ChatGPT/Claude edition. Different products, different instruction fields, different tool surfaces.

When a template is enough instead

If the job is one-shot and reversible — rewrite a paragraph, turn notes into a table — you do not need a fleet. See templates vs fleets. Install multi-agent when the work repeats and some of it can leave the building. If a bot would email a customer, go live, or move money, you need a gate and a named yes, not a sharper template.

What you buy if you do not want to write this yourself

OF-001 is a zip of markdown: CoS, Scout, Product, Ops, optional Research/Sales, routines, gates rulebook, 90-minute sample, templates, first-week checklist. Unzip. Paste into Grok Bot. About 15 minutes to install, 90 minutes for the first run. $29, one-time. Personal / internal use. Checkout is Whop — the button is a placeholder until the live URL is pasted. Support: mike@ownrshp.io.

Related: Chief of Staff template · agent SEND / PUBLISH / PAY gates.