Lieben's engineer runs 20 Grok bots across nine clients
A Chief of Staff bot, one ops bot per client and a 6 PM reporter, with credit caps written into every bot's first message.
Michel Lieben published the structure a senior GTM engineer on his team, Louis, uses to run outbound for multiple clients. It is one Chief of Staff bot plus 20 AI employees, all on Grok, and Lieben says most of it keeps running with the laptop closed.
The Chief of Staff is the only bot Louis talks to directly, often by voice on his phone during walks. It routes asks to the other bots, sends a morning brief covering meetings, tasks due and promises made on calls, and listens to call recordings in Fireflies so it carries client context. Under it sit nine Client Ops bots, one per client, each with its own Slack, Instantly workspace and call history. Those bots pick up requests in Slack, write call summaries and next steps, and recycle silent leads into new campaigns 15 to 45 days later.
Where the work actually happens
List building, copy and scrapers go to Claude Code with the ColdIQ MCP, and the Grok bot checks output quality. Lieben's reason is cost, not capability. The Grok bot can do those jobs itself but it gets expensive in usage, so Claude Code builds and Grok orchestrates.
Replies run on a webhook. An interested reply lands in Instantly, the "Lead is marked as interested" trigger pushes it to that client's reply bot, and the bot posts it in Slack with a company score and a suggested answer for Louis or the client to approve, edit or kill. Each reply bot serves one client only. A reporter bot sends every client a recap at 6 PM each weekday covering emails sent, replies and positive replies, held for Louis's approval first.
The guardrails are the part worth copying
Everything sits on a GitHub repo with a folder per client and a skills folder, which Lieben says to build immediately after the Chief of Staff. Without it, Louis says the output would be "pretty useless".
The rest is damage control. Human approval before anything goes out. A credit cap written into each bot's first message, for example a maximum of 10,000 credits on verifying or enriching emails. Caps on the free lists bots build for prospects who say yes, around 10 companies and 15 verified emails each. Because every bot sees every connected app, each one's first message lists the only connections it is allowed to touch.
The reason for the LinkedIn rule is specific. One of the LinkedIn bots published a post on its own before anyone caught it, so publishing is now set to ask first. For people sending, this is one team's setup rather than a benchmark, and the interesting detail is that the controls are prompt level text in a first message, not platform features.
Early on, one of our LinkedIn bots published a whole post by itself (we caught it & deleted it right away)
Grok Bot can do a lot of these jobs itself, but it's sometimes expensive in usage
Know what is landing before you burn another domain
One email a day. Under three minutes. What the operators found out yesterday.