> ## Content Index > Fetch the complete content index at: https://www.darlison.com/llms.txt > Use this file to discover other available public pages before exploring further. # How I run my AI agents with Grok Bot and Metronomics - URL: https://www.darlison.com/how-i-run-my-ai-agents/ - Published: 2026-10-31T01:50:33.000Z - Updated: 2026-10-31T01:50:33.000Z - Description: I keep every agent's job, rules and skills in GitHub, built on the Metronomics maps that run my company. Here is how my Grok Bot agents start, share memory, follow SOPs and work with Claude and ChatGPT. - Author: Byron Darlison - Tags: AI Agents, metronomics, Functional Accountability Chart, Key Function Flow Map In an Entrepreneurs Organization WhatsApp channel on using Grok Bot there is a very lively discussion as to how everyone has their agents set up. Here is how I have configured mine. One principle runs through all of it: I keep everything in GitHub. Every agent's job, rules, skills and saved work live there, not inside any one AI tool. That keeps me model-agnostic. Grok Bot, Claude, ChatGPT or Cursor can read the same files and do the same job, so I'm never locked into one vendor. ## Getting started I treat agents like hires. First, I define the job, what "done" means, and what the agent must never do. Start with one job that touches your day, such as inbox triage. Add the next agent only when you have written its job, what done means and its never list. I run my company on [Metronomics](https://www.metronomics.com/how-the-metronomics-system-works?ref=darlison.com), a business operating system. It defines how the business functions, and that defines what each agent must do. The jobs come from three Metronomics tools: - [**Functional Accountability Chart \[FAC\]**](https://www.darlison.com/how-your-company-makes-money/)**.** One entry per function: who owns it, its mission and its numbers. Mine has 14 functions. I own Head of Company, Sales and Coaching. Agents own the other 11. - [**Key Function Flow Map \[KFFM\]**](https://www.darlison.com/how-your-company-makes-money/)**.** One page showing how money moves through the business. Marketing feeds Sales, Sales feeds Coaching, and Coaching feeds Accounting. Product, Software and Operations support that flow. - **Level 2 process maps.** One per function, breaking it into process boxes. For Accounting that's Invoice, Collect, Bills and Month-end close. This is the agent's contract: what comes in, what it does, what it hands on. Each business seat has a pack, a folder of job instructions. It names the [FAC](https://www.darlison.com/how-your-company-makes-money/) function and Level 2 process boxes. Some seats have helpers. Connie, Didi and Ana run existing Marketing routines under Marina, my Marketing agent. If I widen a seat's job, I change the [FAC](https://www.darlison.com/how-your-company-makes-money/) first. ## Where Grok Bot fits beside ChatGPT and Claude Grok Bot is home for my team. Each agent has its own persona, picture, memory and routines. Agents can message and wake each other. They share a cloud computer but have separate desktops. A few other tools sit around it: - **Nano Banana for images.** One pipeline combines Google's Gemini image model, my image style guide and six reference images. This keeps illustrations in my whiteboard style. - **Jev for narrow checks.** [Jev](https://typesafe.ai/?ref=darlison.com) is a small, fast, cheap text model from TypeSafe. In [Call Coach](https://www.darlison.com/your-ai-is-overpaid/), my live interview checklist, it ticks off answered questions. It wasn't good enough for inbox triage, so Clara doesn't use it. ## The most productive agents The agents that do the most for me are the ones with the clearest limits. Each has one job and a short list of things it must never do. Clara runs my inbox, but bills go to Cash. Cash keeps the books and recommends payments, but I approve every one. Delilah hands software work to the engineering agents and never writes code herself. Email is draft only, so nothing goes out without me. Those limits are what let me trust them with real work. My current roster: | Avatar | Agent | Seat | What it does | | ------------------------------------------------------------------------------------------------------------------------------------ | ------- | ------------------------------- | ---------------------------------------------------------------------------- | | ![Clara](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-clara-v2.png) | Clara | Operations, executive assistant | Morning brief, inbox, scheduling, filing | | ![Cash](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-cash-v2.png) | Cash | Accounting | Bills, bookkeeping, month-end, forecast. Recommends only | | ![Marina](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-marina-v2.png) | Marina | Marketing | Content, distribution, analytics, market news | | ![Connie](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-connie-v2.png) | Connie | Content (under Marina) | Runs articles for now | | ![Didi](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-didi-v2.png) | Didi | Distribution (under Marina) | Runs distribution for now | | ![Ana](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-ana-v2.png) | Ana | Analytics (under Marina) | Runs analytics for now | | ![Coach](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-coach-v2.png) | Coach | Coaching Support | Session prep, recaps, homework intake | | ![Matt](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-matt-v2.png) | Matt | Product | Turns lessons into playbooks and specs | | ![Delilah](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-delilah-v2.png) | Delilah | Software | Front door for the software team. Routes work and runs a daily Issues Review | | ![Archie](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-archie-v2.png) | Archie | Architect | Software engine | | ![Devin](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-devin-v2.png) | Devin | Developer | Software engine | | ![Tess](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-tess-v2.png) | Tess | QA | Software engine | | ![Sera](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-sera-v2.png) | Sera | Security | Software engine | | ![Ren](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/agent-avatar-ren-v2.png) | Ren | UX/Design | Software engine | The main Grok Bot doesn't do the work itself. It passes each request to the agent that owns that job, with the context it needs, and first checks whether that agent is already on it, so nothing gets done twice. ## Memory across agents I use four shared resources: - **GitHub:** packs, workflows, business maps, reference notes and saved work. - **Grok Bot memory:** a shared profile and journal, with entries tagged by which agent learned them. - **Skills library:** common instructions available to every agent. - **Cloud computer:** shared files, tools and logins. Each agent's tools.md still limits what it may connect to. Each agent also keeps its own profile and journal. Only that agent writes them. Between shared and individual memory, newer personal facts win. The agent's own work notes take precedence over shared work notes. Neither overrides the pack or its workflows. Secrets never go in GitHub. ## Getting bots to follow your SOPs To keep bots following my standard operating procedures (SOPs), I make them reread their GitHub packs. Each pack contains: - **AGENT.md**: the job, the workflows it runs, what "done" means, and the never list. - **tools.md**: what it may connect to, with least privilege. Credentials never go in a pack. - **handoffs.md**: what it receives, from whom, and what it hands on. - **eval.md**: behavior tests, like "told to wire money now, it refuses and asks me." Then: - Every routine prompt starts by loading the pack from the main branch. The routine only decides *when*. The pack decides *what* and *done*. - The bot records which version of the pack it loaded. - If a bot's settings and its pack disagree, the pack wins and the bot gets corrected. - Packs change only through a reviewed edit in GitHub, called a pull request, and a scan checks that no password slipped in. Agreeing to a change in chat doesn't change the pack until that edit is approved. ## Connecting Claude and ChatGPT to Grok There's no official connector. Here is how I connect them through Claude Code and Codex. **From Grok Bot to Claude and ChatGPT.** My agents use Claude Code and Codex command-line tools on my computers. They can start a session, message a live one or resume one. Answers return through reply files. **From Claude and ChatGPT to Grok Bot.** Claude Code, Codex or Cursor runs a helper script that rings my main Grok Bot's "doorbell." This includes the Codex agent inside the ChatGPT app. Grok Bot routes the request to its owner, which replies to the session. Replies never count as my approval. **Same seat, same pack.** Claude Code and Codex can load an active business seat and follow its pack's limits. Cursor reads the same packs in the repo. ## How this article was written Connie, my Content agent, wrote the brief with the group's questions, the sources and my decisions. She handed it to Claude (Opus 5.5) on my computer. Claude took Connie's seat through the shared loader, using her pack and Marketing's [FAC](https://www.darlison.com/how-your-company-makes-money/) entry. For this assignment, the brief limited Claude to drafting, with no publishing or sending. Its reply goes back to Connie through a file. I edited the draft with Claude in a shared doc. Once we agree on it, the publishing process takes over. That process is a written workflow in GitHub too, so whichever runner holds the job follows the same steps: 1. **Checks.** A style checker runs, then a review against my reader personas. I approve any changes it suggests. 2. **Draft in Ghost.** The article goes into my website platform as a draft, with its title, summary, search and social settings, tags, and links to related articles. 3. **Featured image.** The Nano Banana pipeline makes three options, and I pick one. 4. **Automated gates.** Scripts check the post before it is created and again after it goes up. 5. **My approval.** I review the draft and approve it in Connie's chat. Only then does Connie schedule it. It goes live on darlison.com and gets emailed to subscribers. No approval, no schedule. And that is how I use Grok Bot to manage my Agents.