Quick one before we start: Build Night with Hefti is on tonight, 6pm London. Come along.
I built Hefti.
And the bit I want to show you first isn't the mascot, the website or a clever answer from an AI.
It's the test environment.
Yeah, I know. Try putting that on a launch balloon.
But this week we had to finish separating the place where we test changes from the place where people use the product. Different databases. Different payment settings. Different places for a scheduled job to run.
That sounds like plumbing because it is plumbing. And if you're asking an AI system to help run a business, the plumbing deserves a look before the party.
Hefti is now on sale. It's my AI operating system for small businesses. This is why I built it, what it does, and a bit of the less photogenic work behind getting it out the door.
The business shouldn't have to live in your next prompt
I've been building autonomous operations in my own business, with Viktor helping run the back office.
That experience gave me a fairly clear brief for Hefti.
I wanted a small business owner to have AI agents with jobs, a shared understanding of the business, and connections to the tools where the work happens.
So you shouldn't need to introduce your business again whenever you ask for help. Your offers, the way you speak to customers, the things you've already decided. That sort of context belongs with the business.
And a useful answer should lead somewhere. A draft you can review. A recurring job with a result you can check. A tool connection that lets the agent do the authorised task.
You can have a very impressive conversation with an AI and still finish with a desk full of unfinished work. I've built Hefti around getting past that point.
I haven't invented a new AI model. I'm building the working arrangement around the models.
There's a difference, and I don't want to dress it up.
What it does in plain English
Hefti gives you named AI agents that learn about your business and work with a shared business memory.
You can bring in context from ChatGPT, Claude or Gemini. You've already spent time explaining yourself to those tools. Starting again with a blank page would be a fairly annoying welcome present.
But memory needs care as well. An old offer can look perfectly sensible after you've stopped selling it. An imported instruction can clash with how you work now.
So I treat bringing information across as something you review, not proof that everything the old tool knew still applies.
Then you've got the jobs themselves.
Hefti connects to supported apps, lets you schedule work, and gives you access through the app, Slack or Telegram. The point is to work where you already are, with the tools you've connected.
That last bit matters. An agent can't inspect a calendar you've never connected. Giving it a friendly name doesn't grant it access.
And you still need to decide what it may do.
A useful brief might be, “Prepare the meeting notes and show me the draft.” That's quite different from, “Send whatever you think is appropriate to everyone involved.”
I'd start with the first one.
Once you've seen the results and checked the permissions, you can decide whether the next step makes sense. There's no prize for handing over the riskiest job first.
You're not marrying a model
Hefti works across models from different providers.
You can let Auto choose a model for the request, or pin a favourite. A short routine task and a difficult piece of reasoning don't necessarily need the same treatment.
But I won't tell you the cheaper answer always matches the stronger one. Answer quality varies. Routing needs testing against the sort of work you want done.
That's why I care about the result, not just the model name sitting above the chat.
If you're using it to prepare a customer-facing draft, read the draft. If you're asking it to organise information, check what it kept and what it left out.
And if a model provider changes something, your business context shouldn't depend on your loyalty to that provider's logo.
That's the design choice. The business keeps its working context while the model choices can change.
The bit that needed fixing before more people arrived
Back to that deeply exciting test environment.
During the move to the live site, we still had the same database behind what we'd been calling staging and what had become production.
The names suggested more separation than we had.
So we split them.
Production now has its own database. Staging has a separate one, with the database structure copied across without the customer data. Staging uses test payments, and its scheduled work points at the staging system.
We separated the Vercel preview settings as well. A preview is a temporary version of the app for checking changes. It shouldn't get the live database simply because someone copied the production settings and forgot about them.
Look, none of this makes a dramatic demo.
You click a button and the correct, separate database receives the request. Nobody applauds. That's probably a healthy response.
But the alternative is testing a change against the people who signed up to use your product.
And I don't think calling something a soft launch makes that a sensible testing plan.
Some third-party services still share accounts across the environments. Separate databases don't magically separate every supplier. We have to keep checking the connections as well.
I'm telling you that because “we split staging” should mean something specific, not “everything is now perfect”.
We fixed a gap in the import history as well
Another recent repair involved the record of what happened when information came into Hefti.
The import audit needed to account for skills and channels. Skills are the instructions for how an agent does a job. Channels are the places conversations happen.
So we added those missing audit categories, taking the database change through staging before production.
Why bother with that when you could be making the chat box prettier?
Because if someone asks what came across during an import, I'd like the system to have a useful record.
And when something looks wrong later, that history helps you work out where to look. Did the imported instructions contain it? Did something change afterwards? Without a record, you're left reading today's result and guessing about yesterday.
That repair doesn't mean an audit history can explain every AI answer. It can't.
But it gives us a better starting point than a shrug.
This is the sort of thing building the product keeps putting in front of us. You make the visible part work, then check the records and permissions around it, and find another bit that needs attention.
What I wouldn't hand over on day one
If you're trying a system like this, I'd pick a recurring job you can inspect without much effort.
For argument's sake, take an internal meeting brief.
Tell the agent where the information lives, what the brief should contain, and where you want to review it. Connect only what that job needs. Keep the first result away from customers.
Then read it.
Did it use the current information? Did it admit what it couldn't reach? Did it produce the thing you asked for, or a lovely explanation of how someone might produce it?
That last one gets old pretty quickly.
And if it makes a mistake, fix the working instructions or the source information before scheduling more of the same mistake.
You can apply that test whether you use Hefti or something else. A short, checkable job will teach you more than telling an agent to “sort out the business” and hoping it shares your definition.
It's a soft launch, so I'm showing the thing
Hefti has a seven-day free trial. I'm its founder, and I want you to know that when I talk about it.
I'm not going to pretend every idea on the roadmap already works. The browser agent and Spaces app builder, for example, sit under “Coming next” on the site.
What I can show you is the product we have now, the jobs it can do with the connections available, and the decisions we're making as we build it.
And I'll show you where I still want it to improve.
I've spent plenty of time talking about businesses that run themselves. Building Hefti means putting more of those ideas into something other people can use and judge.
So tonight I'll open it up on screen.
No point hiding behind the launch page.
Sources
- Hefti product page and current feature availability, checked 11 November 2026.
- Build details come from our own release records for the staging separation and import-audit repair on 9 and 10 November 2026.
Tonight's Build Night with Hefti is at 6pm London, live on Zoom. I'll demo Hefti itself, show you how it works and talk through the bits behind it. If something misbehaves on camera, you'll see that as well. Grab a seat.
Brewed by Steven, poured by Hefti
About Steven Tann: Steven helps business owners build systems that run themselves using AI. After 10+ years helping 7,000+ businesses and building his own autonomous operations, he's the bloke who actually does it, not just talks about it. He's also the founder of Hefti, the AI operating system for small business. Find out more at steventann.com.