AI agents in your code: less magic, more control

You asked for a new screen in your system. The vendor said "AI can do this in an afternoon". Three weeks later, nobody can explain why the sales report stopped adding up. This story is becoming routine. AI agents write code at an absurd speed. The problem is almost never speed. It's what happens when nobody is holding the wheel.

A few days ago a project called Huzzah showed up on Hacker News. The author presents it as a new way to code with AI. I won't summarize the tool here, since it's meant for developers. What caught my attention was the question behind it, and it matters a lot to anyone paying for software: who's in charge of the work, the person or the agent?

What AI agents are in software development

A chatbot answers questions. An agent acts. It opens files, writes code, runs tests, fixes the error and tries again. All on its own, one step after another.

In practice, the developer writes something like "create a customer registry with search by tax ID" and the agent gets to work. In 20 minutes there's a working screen. It looks like magic. And that's where the danger lies.

Because the agent doesn't know your business. It doesn't know your finance team needs the tax ID formatted with punctuation for the old ERP. It doesn't know inactive customers can't show up in the sales team's search. It fills those gaps with reasonable guesses. And in a company's software, a reasonable guess is a bug that looks good.

Why the "ask and pray" model goes wrong

The most common way to use AI for coding today is what I call ask and pray. You describe what you want in a long text, the agent generates a pile of code and you hope it's right.

It works well on small projects. A landing page, a script that organizes a spreadsheet, a prototype to show your business partner. In those cases, if it goes wrong, you throw it away and ask again.

In a system that runs your operation, the math changes. Here's an example I saw up close. A company asked for an automation to generate invoices from approved orders. The agent delivered fast. But it read "approved order" as any order with a status other than "canceled". The result: invoices issued for orders still under credit review. About 40 wrong charges went out before anyone noticed. The code was clean. The rule was wrong.

AI doesn't fail for lack of capability. It fails for lack of context.

What changes when the person stays in control

The idea that projects like Huzzah explore, and that I've defended for a while, is simple. Instead of handing the whole task to the agent, you break the work into small decisions and review each one.

It seems slower. It isn't. Compare two paths to the same feature:

• Ask and pray: 1 hour generating, 6 hours figuring out what went wrong, 4 hours fixing it without breaking everything else.
• Step by step with review: 3 hours building in short stages, 1 hour of fine tuning.

The numbers vary from project to project, of course. But the pattern repeats. The time you "save" by letting the agent run free comes back with interest during the fixing phase.

Step by step, the flow looks roughly like this:

1. The person defines the business rule in plain language, before any code.
2. The agent proposes a short plan. The person approves or corrects it.
3. The agent implements a small part.
4. The person reviews it, tests it with a real case and only then releases the next part.

It's the same way a good manager works with a new hire who is very capable and has zero context about the company. You don't hand over the keys to the safe on day one. You give tasks, check them, adjust and loosen the reins as trust grows.

How to tell if your vendor uses AI the right way

You don't need to understand code to ask good questions. And today practically every developer uses AI in some way. What matters is how.

Some questions that separate those who use it well from those who just push the button:

• "Who reviews the code the AI generates?" If the answer is "the AI itself", that's a yellow flag.
• "How do you validate my business rules?" Expect to hear about test cases with real examples from your operation.
• "If something breaks six months from now, will anyone understand this code?" Code nobody has read is code nobody knows how to fix.
• "Are deliveries big or in parts?" Small deliveries are a sign of control. One big package at the end of the month is a sign of praying.

My opinion, no sugarcoating: a vendor selling "AI does everything, that's why it's cheap" is selling you debt with installments that come due later. The real savings from AI come from clearing the grunt work out of the way, not from removing human judgment from the process.

Where AI agents truly shine

I don't want you to leave thinking AI agents are a scam. Quite the opposite. I use them every day, and the gain is huge when the work is well divided.

They're great at:

• Repetitive tasks: creating ten similar registration screens, migrating data from one format to another.
• First drafts: getting an idea off the ground in hours so you can see it, click around and give feedback.
• Reading legacy systems: explaining what code from 2015 does, something that used to cost someone days of suffering.
• Tests: generating test cases a tired human would forget.

And they're weak at anything that depends on knowing how your company really works. That exception only the warehouse team knows about. The big client with different payment terms. The workaround in the tax system nobody documented, but everyone respects.

None of that comes from a model. It comes from conversation. That's why the most valuable part of an AI project is still the most human one: sitting down, understanding the process and writing the rules before generating any code. If the agent were an intern, it would be the fastest intern in the world. Still an intern.

What your company should take from this

If you're thinking about using AI to build or improve a system, keep three ideas in mind:

Generation speed isn't delivery speed. What counts is how long it takes until the thing works right, in production, with your data.

Control means working in small parts. The bigger the chunk you hand the agent at once, the bigger the chance a wrong assumption is hiding inside it.

Context is your asset. Knowing how your business operates is worth more than any tool. AI amplifies that knowledge. Without it, AI amplifies guesswork.

The debate Huzzah sparked among developers is the same one every business owner should have with whoever handles their technology. It's not about using AI or not. It's about who decides.

That's how I work on my automation and AI projects: the machine speeds things up, I make the decisions and you follow every step. If you want to understand this way of working better, learn a bit more about me.

The vendor said "AI can do this in an afternoon". Three weeks later, nobody could explain why the sales report stopped adding up.

I once saw an AI agent issue 40 wrong invoices with flawless code. The code was clean. The business rule was wrong.

AI doesn't fail for lack of capability. It fails because it doesn't know your business.

Asking for everything at once and hoping it works feels fast, but the bill comes when it's time to fix things. What works is splitting the work into small parts, reviewing each one and only moving to the next after testing.

The machine speeds things up. People still make the calls.

If you want to use AI in your systems without becoming hostage to it, let's talk.

#ArtificialIntelligence #AIAgents #SoftwareDevelopment #Automation #Technology