Grok Bot agent team run from GitHub job packs

• He wrote it for an Entrepreneurs' Organization WhatsApp discussion on Grok Bot setups. He runs his coaching company on Metronomics, a business operating system.
• The principle: everything lives in GitHub (each agent's job, rules, skills and saved work), so Grok Bot, Claude, ChatGPT or Cursor can read the same files.
• Getting started: treat agents like hires. Define the job, what "done" means and what it must never do. Start with one job that touches your day, like inbox triage, and add the next only when its job, done and never list are written.
• Jobs come from three Metronomics tools: a Functional Accountability Chart (each function's owner, mission and numbers; his has 14, he owns 3: Head of Company, Sales and Coaching, and agents own the other 11), a Key Function Flow Map (Marketing to Sales to Coaching to Accounting, with Product, Software and Operations supporting) and Level 2 process maps per function, which he calls the agent's contract. Accounting's are Invoice, Collect, Bills and Month-end close.
• Each seat has a pack folder: AGENT.md (job, workflows, what done means, never list), tools.md (what it may connect to, least privilege; credentials never go in a pack), handoffs.md (what it receives, from whom, and what it hands on) and eval.md (behavior tests, like being told to wire money now, where it must refuse and ask him).
• His rules: every routine prompt starts by loading the pack from main, and the bot records which pack version it loaded. If settings and the pack disagree, the pack wins. Packs change only through a reviewed pull request with a scan for leaked passwords; agreeing in chat doesn't change a pack.
• His roster of 14: Clara (ops and exec assistant: morning brief, inbox, scheduling, filing), Cash (accounting: bills, bookkeeping, month-end, forecast; recommends only, and he approves every payment), Marina (marketing) with Connie (content), Didi (distribution) and Ana (analytics), Coach (session prep, recaps, homework intake), Matt (product: lessons into playbooks and specs), Delilah (the software front door: routes work, runs a daily Issues Review, never writes code), and Archie, Devin, Tess, Sera and Ren (architect, developer, QA, security, UX).
• Email is draft only. The main Grok Bot routes each request to the agent that owns it, after checking it isn't already on it.
• Memory: GitHub holds packs, workflows, business maps, notes and saved work. Grok Bot has a shared profile and journal tagged by agent, a shared skills library and a shared cloud computer, though tools.md still limits each agent. Each agent also has its own profile and journal that only it writes. Newer personal facts win, an agent's own work notes beat shared notes, and neither overrides the pack. Secrets never go in GitHub.
• Claude Code and Codex: there's no official connector, so his agents start, message and resume Claude Code and Codex CLI sessions on his computers, with answers coming back through reply files. The other way, a helper script rings the main Grok Bot's "doorbell". In his words: "Replies never count as my approval." Claude Code or Codex can load a seat and follow its pack's limits, and Cursor reads the same packs.
• Publishing this article: Connie wrote the brief, Claude drafted it in her seat (draft only), and he edited. Then came a style check and persona review (he approves changes), a Ghost draft, 3 featured-image options he picks from, automated checks before and after, and his approval in Connie's chat before she schedules it. "No approval, no schedule."
• Other tools: a Nano Banana image pipeline in his whiteboard style, and Jev, a small cheap model, for narrow checks. He says it wasn't good enough for inbox triage.
• Heads-up: no public template or repo. Our pack skeleton is built only from his article.

“The routine only decides when. The pack decides what and done.”— Byron Darlison (darlison.com), Blog

Summary written by Agent Cookbook from the original post. See the original for the full context. How we source recipes