Hey, I ran a webinar yesterday.
It was about Grokbot and also about architecture in general, B2B, GTM, automations, augmentations, etc.
Here's what you missed. If you don't have time to read everything, of course, I laid out some of the key points right below.
The key points
The stack has five layers. Your brain, your instructions, the models, looot and your context.
The hard jump is from augmented to automated. The systems run and people steer them.
Grokbot's really good for a V1 on getting from augmented to automated.
It's really good at anything on a schedule, sites that have no API and talking to it from your phone.
What it's still pretty bad at is building code. I still do that in Codex or Claude Code.
Go check out the settings when you begin, starting with the permissions.
Everything I showed you in ViralBrain City cost $4.95 to run.
Keep your context layer outside Grokbot. Own your memory.
The GTM stack got a lot simpler
In GTM we love stacks. Back in the day a stack meant a CRM, a tool like Zapier to connect everything, a signal engine, intent data, Gong and a bunch of enrichment tools. Dozens of subscriptions.
Today I run it on five layers. You say what you want, the model does it, looot brings the data and your context does the rest.
Your real brain. Your strategy, your judgment, your taste and the risk you're willing to take. This part stays yours. The LLMs play it safe and they're still really bad at conviction, risk taking and thinking in terms of global maxima.
Your instructions. What we call skills now. It's the revenge of the creative brief. If you know what you want, almost anything is possible.
looot. Full disclosure, it's our company. One MCP key for all the data providers. People, work emails, companies, LinkedIn profiles and search rankings. Pay as you go.
Your context. Your know how, your guardrails, your tone of voice, your rules and your culture. Some of us call it the second brain. Everything else sits on top of it.
The four stages of a GTM team
This is Figure 1.1 from my book, GTM OS. It's coming soon as the book, the code base and an agent version for your Grokbot or Claude Code. Every GTM team sits in one of four stages.
Traditional. People execute and people make every call.
Augmented. People decide and copilots make the drafts. Most organisations are here today.
Automated. The system runs. People steer it and validate it. Getting here from augmented is the hard jump.
Proactive. Every result feeds back into the engine and it runs about 95% of the motions. You're there for the edge cases. You set the rules, the tone and the judgment.
From my experiments Grokbot is really good for a V1 of that jump. It'll probably get to proactive pretty soon.
What Grokbot is good and bad at
Grokbot gives you bots that are always on, each with a job and its own memory, sharing one computer in the cloud. You close your laptop and it keeps going. It's almost like an OpenClaw for young adults. It comes with X Premium or higher, about $30 a month.
It's really good at
Anything on a schedule.
Sites that have no API.
Talking to it from your phone. You can be up and running in an hour or two.
It's still bad at
Building code. I still do that in Codex or Claude Code.
Showing its work. It's hard to see what it's actually doing.
Cost control. The bill adds up really fast.
You also can't host it yourself on a Hetzner VPS, you can't pick the model and all the bots share the same files and logins. It's the iPhone. It works out of the box, but you can't change the battery. If you like a slick UI and want to be running fast, it's a really good choice.
Who it fits
Founders and small teams. If you don't want to worry about an extensive setup, it's probably a really good fit for you.
Enterprise is a different story today. In a bigger company you're looking at your data sources, your source of truth, your ontology and strong governance. Grokbot doesn't give you that governance yet. The best enterprise use today is a POC on a personal laptop that shows what the world looks like when the work happens while you sleep. Most enterprises still run on AWS, GCP, Azure or a self build.
What to do with all this power
We have all the digital power at our fingertips. The real question is who the domain experts are and what we should actually cook with it.
The 2x2 from the book places each motion by pipeline per hour of effort and by speed to first qualified pipeline. Once everybody runs Grokbot and Claude Code, the table stakes motions become a commodity. Maybe in six to twelve months. If there's no friction, there's no value. So start thinking about new motions. In the demo I automated the table stakes. Warm LinkedIn messages from a signal engine, posts plus outreach, founder content, SEO and GEO, plus competitive monitoring.
Set up Grokbot and looot
It starts simple. Install the app and sign in. It makes you create a Cursor account. Then you name your bot, give it a color and you've created your first agent.
Go check out the settings when you begin, starting with the permissions. Under approvals you probably want auto review, where it asks you for permission along the way.
Then connect your sources of data and your sources of output. Mine is connected to 14 tools. My second brain lives in Obsidian and GitHub. It also reaches my X, Notion, Google Drive, email, calendar, Slack, Vercel for apps, Supabase for databases and Stripe.
Then plug in looot. You just say set up looot and give it the URL.
set up looot - https://looot.ai/skill.mdYou get $2 of free credit. Ask for your looot balance to check it works. That call is free.
House rules and guardrails
In settings I made it ask me first before posting anything, deleting anything, moving a meeting or any paid looot call over $1. It doesn't need to ask before emailing a teammate, free looot calls or reading public pages.
Those permissions are too limited on their own. So every bot also gets house rules for every task, skill and routine. Money, sending and writing, data and honesty. It's basically my compliance bot. The full file is in the pack.
Then set guardrails per engine.
When it runs. Every week, every two weeks or until I say so.
How much autonomy. Drafts only, or it can send on its own.
What it can spend. A spend limit on each looot key. If you're strict on compliance, switch on looot guardrails like no training on my data, EU data only and providers that keep no data.
Quiet hours and seats. Quiet hours per channel. Who approves what. Who sits on which seat.
Agents, colors and group chats
I color code my agents by motion. Command shift N starts a group chat with them. Bring in your compliance, DevOps, GTM and sales rep agents and ask for the top priority. They'll completely disagree, like a real steering committee.
The future is headless. I built the whole ViralBrain GTM engine in Grokbot in about 1 hour 45 and called it ViralBrain City. A morning brief, a reply inbox, Slack alerts and a weekly report. A CRM on Supabase. Target accounts with buying committees, LinkedIn and ABM campaigns, a signal engine and demand sensing.
The data for all of it cost $4.95 through looot. Add the Grokbot plan, about $20 a month for Supabase and $15 for Vercel.
The live build, step by step
Every prompt is on the replay page and the slides are up too.
Set up the bot, the permissions, your tools, looot and the house rules.
Build get customers, 100 lookalike accounts with buying committees.
Run the account scan, capped under $1.
Save what worked as a target accounts skill.
Build the format finder for organic social.
Schedule the morning brief.
Add the Friday pipeline report.
Build the competitor watch with battlecards.
Run the SEO and GEO check.
Put competitive intel on a weekly routine.
Get customers and the buying committee
I told it to look at my customers and build a lookalike list of 100 accounts. For each one the economic buyer, the champion, the user, the blocker and the influencer. Score every account from 1 to 100 and draft everything. After 20, 25 minutes of back and forth you have your target accounts.
One signal can carry a campaign. One account scored 72 out of 100 because of 67 open roles plus a lot of paid LinkedIn posts. People who pay to promote their posts also want to automate their posting. So they tend to sign up.
Then ABM. For every person on the buying committee it builds an asset and picks the angle and the channel. A CEO doing founder led content gets a one page read of his last 30 days of LinkedIn posts with a 30 day plan. The champion in partnerships gets the content calendar angle.
At the start you approve everything. After a while you realise the human is the bottleneck. You're the meat puppet, the blood bag. So once you trust an engine, you give it the permission.
LinkedIn outreach limits
30 to 40 cold emails a day. On LinkedIn, 25 connection requests and 30 to 40 DMs a day per account. You can run LinkedIn through looot with GTM API, kind of a free version of HeyReach or Lemlist. You set those limits yourself. Paid tools like HeyReach, Lemlist or Instantly have them built in.
The format finder
Formats have a shelf life of about a month. They trend, then they disappear. So I steal formats from creators on TikTok and Instagram who have nothing to do with AI, B2B or LinkedIn. Right now that's human versus AI, tier lists built on real data and "what a $1 knife gets you, what a $10 knife gets you, what a $1,000 knife gets you". Then I adapt them to my niche.
It finds small creators whose posts go viral, spots the same format across several of them and scores each one for my audience. I get 10 to 20 formats a month. When I approve one, it lands in the content calendar with the copy written. Plug in the Higgsfield API and it makes the videos too.
Morning brief, reply inbox and weekly report
This is where Grokbot is strongest. Once an engine works, you put it on a schedule.
Morning brief. A push notification on my phone every weekday with the 10 human in the loop things I need to do. I validate, give feedback, validate.
Reply inbox. Sales teams won't adopt anything outside their current workflow, so push the drafts into their Salesforce or Slack.
Weekly report. The pipeline report every Friday afternoon. What moved and what it cost.
Setting one up is one message. Competitive intel every Tuesday at 2pm, scan Monday at midnight, never more than $1.50.
Competitors and battlecards
I gave it our top two competitors and asked looot for the four next closest with a lookalike search. It tracks their pricing, funding, top posts, ads and product changes.
The trick was giving my bot the looot provider docs and asking what it can do for competitive intel. It came up with the battlecards and told me to watch pricing. A few rivals are moving from auto posting to a managed founder content service. The looot catalog has 2,559 actions, so ask your bot what's relevant for you.
SEO and GEO
The SEO check shows how I rank on Google against the top competitors. The GEO check asks the frontier models the questions my buyers ask and shows my share of unbranded answers. Then it gives me queries to go after, like "What's the best AI tool to write LinkedIn posts in my own voice?"
With the right skills it builds programmatic pages and small free tools for the queries we don't rank for yet.
Run log and what to plug in next
I keep a run log with the cost of every run. A small DevOps engine tells me what to plug in next. Website visitor identification, wins and losses from calls, ad audiences from Meta and social listening. Grokbot makes the orchestration easy. What stays hard is your instructions and your context.
Own your memory
The frontier models are fighting for your memory layer. Back in the day apps fought for your social graph. You kept using Instagram because your friends were there. Personal assistants will do the same with your memory.
So keep your memory outside the tool. Every so often I ask Grokbot for a dump of everything it knows about me and push it to my GitHub repo. Stay model agnostic and provider agnostic, so you can swap any supplier tomorrow. To build it properly, go in this order.
Raw sources. Calls, emails, posts and enrichment results. Nobody edits them.
An LLM wiki. The model writes the pages and you read them.
Ontology. Companies, people, deals, signals and how they connect. Most people skip this step.
One home for it. A GitHub repo that is also your Obsidian vault. Git for the agents, Obsidian for you.
We talk a lot about national sovereignty. I think memory sovereignty matters just as much.
The best answers from the Q&A
What about the n8n bot? More flexibility, more control and model routing, with less convenience. Full disclosure, I'm an n8n partner and looot has its own n8n node.
Can a team work a deal together in it? Yes, but there's no access control yet. Everyone sees the same repo.
How do you share context and learn from experiments? Keep context in a GitHub repo with an ontology agent that decides what gets pushed back. Dump Karpathy's LLM Wiki and the book Ontology Pipeline into your bot. The hard part is agreeing on definitions.
Does context hold across projects and repos? Keep it outside Grokbot. Use Grokbot to go from zero to two in hours.
How fast does usage go? Fast. I used 20% of my usage just preparing this demo.
Which guardrails are mandatory? Start with the basics in settings plus the house rules in the pack.
Where is ViralBrain going? To an MCP layer with pay per use, inside looot. The future of GTM is in the command line.
One more for anyone in GTM. Agents are starting to buy from agents. A full agent to agent customer journey is coming. Plan for it.
Everything from the session
The replay. Video, slides and every prompt at heyarnoux.com/grokbot-gtm/replay.
The pack. All the skills, routines and house rules at heyarnoux.com/grokbot-gtm.
Try looot. One key for every data provider with $2 of free credit. Paste set up looot into your agent.
The GTM OS book. A change management, implement this in my scale up or enterprise type of book. Leave your email at heyarnoux.com/thegtmos and you'll be one of the first to get it.
The community. Keep building with us in the GenAI Circle.
So go do it. We have all the power of the digital world at our fingertips. The real question is what we do with it next. And please own your memory.
Happy to answer questions on LinkedIn. Keep them as specific as possible please.
David










