01What a Long Running AI Agent Is and How Carbon Fits
A long running AI agent is an AI that keeps working on one job for hours, all by itself, and checks its own work as it goes. Gemini 4 Carbon is Google's new model built for exactly that kind of job.
In this episode, Julian Goldie puts that claim to the test. Instead of covering the news again, he gives Carbon two jobs that AI Profit Boardroom members keep asking about and lets it run with one prompt each.
- A client booking system, so you stop renting one.
- A follow-up desk for your leads, with review requests built in.
The thing that makes Carbon different is stamina. Google says it keeps working on a task for a very long time and checks its own work along the way.
The big question Julian asks is simple. Can it build something a real business would use while nobody watches? The answer is yes, with the right setup.
That setup is four rules, and he says the prompt matters more than the model.
02Where You Can Get Gemini 4 Carbon
Google announced Gemini 4 Carbon on the 28th of October. At first it was for cybersecurity teams only.
On the 10th of November, it opened up to more people. You can now get it in these places:
- The Gemini API
- AI Studio
- Gemini for Business
- The Gemini app and Antigravity, if you are on Google AI Ultra
For long builds like the ones in this episode, the setup starts in AI Studio or Antigravity. You give the model a project folder to work in, which matters for the rules you will see below.
If you only want to try Carbon for a quick chat, the Gemini app is the simplest door. But to use it as a long running agent that builds a full app, you want a setup where it can open files, save progress and run a browser to check its own work. That is what the rest of this article covers.
03The Four Rules for a Long Running AI Agent
Julian Goldie shares four rules that make long runs work. These go into the prompt before the build starts.
- Set a minimum number of passes. Long-run models will happily call something finished after pass five. He told Carbon to run at least three hours and at least 60 passes, with one small improvement each pass.
- Give it eyes. Every pass, Carbon opens the app in a headless browser, takes a screenshot and reads the console for errors. Every 10 passes, a fresh reviewer with no memory of the build looks at the app and writes a list of what is wrong.
- Keep everything that matters in files. One file is a checklist of what done looks like. Another is a progress log with one line per pass. If the run crashes, it reads the files and picks up where it left off.
- Check the result yourself. Check in on the run and test what it built with your own hands.
The progress log has a second benefit. Afterwards you can read exactly what happened, line by line.
04Build One: A Client Booking and Intake App
The first job was a booking and client intake app for a coaching and consulting business. One prompt. About an hour and a half later, here is what Carbon built.
On the client side:
- A premium homepage with a 3D orb and live numbers.
- You pick your advisor and the kind of session, and a week view shows every open slot.
- Pick a slot, and an intake form asks the right questions before the call. It checks every field and saves your draft as you type.
- Book it, and you get a confirmation with a reference code, a calendar file and a prep checklist.
On the coach's side, the advisor desk shows every upcoming call. You can search, filter, open a client's full brief, reschedule, cancel, mark a call done, block out time and export the list.
It works on a phone too. Julian Goldie points out that this is a real client system, and nobody wrote a line of the code.
05What the Fresh Reviewer Found During the Run
Rule two paid off. Every 10th pass, a fresh reviewer checked the app with no memory of the build. It found 12 real problems.
A few examples:
- The page scrolled sideways on a phone.
- Keyboard focus escaped the pop-ups.
- A score said 100% when there were no calls.
Every one of those went straight onto the checklist. Carbon had already fixed the phone layout when Julian Goldie ended the run, and the rest were next in line.
Then there is the honest part, which is rule four. Twice the run got stuck. Once, Carbon saved a screenshot outside the project folder.
Antigravity stopped to ask about it, and the run waited until Julian checked in. He said no, pointed it back to the right folder, and it carried on.
So yes, a long running AI agent like Carbon builds on its own. But you still need to check in on it and check the result yourself. A reviewer with no memory catches the small bugs that the builder stops seeing.
06Build Two: Follow Up Desk for Leads and Reviews
The second build came from a member's post. A web designer wanted to sell clients a proper system for leads, follow-ups and reviews, instead of renting one. So Julian Goldie asked Carbon for exactly that.
It is called Follow Up Desk. About half an hour and 18 passes later, here is what it had:
- A dashboard with live numbers and a chart that draws itself.
- A leads inbox where every lead gets a score.
- A pipeline that runs from new to won.
- Open a lead, and it drafts the next three follow-up messages for you.
- When a deal is won, it sends a review request with a preview of how the review will look.
This one is a good fit for agencies and web designers. It is a tool you can build once and sell to many clients.
Notice the time difference. The booking app ran for about an hour and a half. Follow Up Desk ran for about half an hour.
Bigger jobs simply need more passes.
07How to Set Up Your Own Long Running AI Agent Build
Here is the setup Julian Goldie walks through, step by step.
- Open AI Studio or Antigravity and pick Gemini 4 Carbon.
- Paste a long-run prompt with the goal, the checklist and the four rules: a minimum number of passes, a screenshot and a fresh review, state kept in files, and you checking the result yourself.
- Give it a project folder to work in.
- Set permissions so it never stops to ask, then let it run. Before bed is a great time to start one.
- When you come back, open the progress log and the last screenshot first.
- Then test the app yourself.
Step four matters because of what happened in build one. When Antigravity stopped to ask a question, the whole run sat waiting. Setting permissions up front avoids that.
Step five saves you time. The log tells you what changed on every pass, so you know where to look before you start clicking around.
08What to Leave Carbon Running On Tonight
Julian Goldie gives a simple rule of thumb. A good job for Carbon is anything you could describe to a developer in one page, with a clear checklist of what done means.
Ideas from the episode:
- An agency could build a client report dashboard.
- A course owner could get a student onboarding portal.
- An ecommerce store could get a product finder quiz.
- A consultant could get the booking app shown above.
A few mistakes to avoid:
- Do not let it stop early. Set a minimum number of passes.
- Do not skip the screenshot and console check. Without eyes, it cannot see its own bugs.
- Do not rely on memory alone. Keep the checklist and log in files.
- Do not trust the result blind. Test it yourself.
This is for business owners who want their own tools instead of renting them, and who are happy to write one clear page describing what they want.