What four people at Hostinger actually do with AI all day (and what happens when you have an agentic beef)

An AI engineer, a software engineer, an influencer partnerships team lead, and an operations manager at Hostinger walk into a bar.

Between them they’ve got fleets of custom-built AI agents, automated workflows that replace hours of manual work, and personal prompt libraries they’re a bit too obsessive over.

Tasks that used to take half a day now take minutes. Processes that needed a whole team run in a few clicks.

They’re shipping faster, debugging smarter, and finding patterns in data that would have taken weeks to spot manually.

Hence the bar. The agents handle the work, they take care of the rest.

Sounds like they’ve got it all figured out, right? Not even close.

Agents skip steps, suggest partners that have already been contacted, hallucinate details about YouTube channels, and propose code fixes that would definitely take down a whole system.

One of the four is even in a full-on beef with an AI agent that’s developed an attitude.

It’s beautiful chaos. Here’s what it all actually looks like.

The things they’ve built

Eylül Güleryüz is an AI engineer, which is a job title that didn’t even exist a few years ago. It sounds like a fancy title from the future that just showed up early.

And it kind of is. She has 10 agents running in parallel on any given day, each on a different task. Her job used to be 80% her and 20% AI. That ratio has completely flipped. Now, she’s a “manager of agents.”

“My entire development process is automated,” she says.

“I don’t even open Jira to pick a task anymore. I have AI workflows that figure out the most impactful things to work on, suggest where to start, and then we just… go. I confirm the plan, and from there it handles everything all the way to having a pull request ready for me to review.”

What does all this sorcery actually involve?

She calls it “harness engineering” – building the systems that let LLMs perform tasks in the smartest and most efficient way possible.

“Most of my time goes into creating the right environment for the models, testing everything thoroughly, and figuring out new ways to push the results further. It’s a lot more engineering than magic words.”

Someone who’s also running a fleet of agents is Karolis Varnelis. He leads influencer partnerships expansion into new markets, which means every market is different, every audience is different, and there’s no single playbook that works everywhere.

So he built one. Well, several.

“I have a lead sourcing agent that helps find potential influencer partners. A team performance analysis agent that highlights key insights. A partner performance analysis agent that flags when performance is dropping. A CRM audit agent that keeps our internal CRM clean and organized. And a content quality agent that checks whether YouTube integrations include all the key details.”

Five agents, each with a specific job. That’s someone who’s moved well beyond dabbling and rebuilt how their entire function operates.

Gabrielė Bagdonaitė is the influencer performance and operations manager, and her specialty is making painful manual processes disappear.

Exhibit A: the team’s invoice generation flow. A year ago, they were processing invoices from over 30 stakeholders by hand. One to three hours a day. It was as painful as it sounds. Now it’s a few actions in an AI-based workflow.

She’s also built agents for data checks, meeting recaps, and partner reviews, and uses AI to search internal conversations for precedent.

“Instead of spending forever digging through messages, I can basically ask ‘have we dealt with this before?’ and get to the right context much faster.”

Then there’s Ugnius Šmaižys. He’s a software engineer, and AI has reshaped his day just as much.

“Debugging is probably the biggest change for me,” he says.

“Before, I could spend hours going through Grafana/Loki logs, following correlation IDs, checking graphs for anomalies, and tracking down exceptions. Now I can give AI a lot of that context and let it connect the dots. Something that could take me a few hours to track down can sometimes take just a few minutes now.”

But Ugnius hasn’t built a fleet of agents or automated a whole pipeline. He’s built a deep enough understanding of the systems he works on that he knows exactly where AI is useful and where it’s a liability. That context is his real advantage.

“People probably think the hard part is writing the code, but usually the harder part is understanding the system well enough to know what code you should write.”

Four very different approaches to the same idea: figure out where AI is useful, build around it, and keep going. Of course, it also argues back, skips steps, and occasionally tries to burn the house down.

The beef (and other things that break)

“I actually have a ‘situation’ with one of my AI agents right now. We definitely have our hallucination moments and our little beefs with each other.” That’s Gabrielė.

She built an AI agent to handle partner content review – scraping a partner’s video, checking links and coupons, verifying the product is represented correctly, and logging it all in a sheet.

Sounds straightforward until you count the systems involved: YouTube, HubSpot, coupon tools, internal platforms, different spreadsheet tabs.

“There are so many variables that sometimes it just gets confused,” she says. “It completes part of the task but misses another part, or suddenly tells me it doesn’t have permission to access a system when I KNOW it does. So we have conversations like, ‘You do have access.’ ‘I don’t have access.’ ‘Yes, you do.’ And then suddenly it finds the tool and continues.”

It’s like working with someone who has selective amnesia but unlimited confidence.

Is she throwing it out? No. She’s breaking the workflow into smaller pieces, restructuring the logic, and simplifying things.

“Sometimes the problem isn’t that AI can’t do something. The workflow might simply be too complex, or the logic isn’t structured in the right way for the agent yet.”

Eylül knows this better than anyone. When an agent gets confused about permissions or skips a step, it often comes down to the harness around it, which is the context, the instructions, or the constraints.

That’s where a lot of her time goes, tweaking the harness until the agents do what they’re supposed to.

Karolis has his own version of an agentic beef. His lead sourcing agent keeps suggesting partners already in their system, even when told not to.

“Quite often, AI provides leads that have already been shared before, even when you clearly tell it not to,” he says. “You always need to double-check, because mistakes or incorrect information can still happen.”

And Ugnius has caught something more dangerous.

“AI can struggle with systems where small details have a lot of consequences. I’ve had cases where it correctly found the cause of a problem and suggested a fix that looked completely reasonable, but because I knew the system well, I could see that one small part of the change could break something much bigger.”

The code was technically correct in isolation. It would have been very wrong in context. Good luck getting an AI to spot that one.

The stuff they keep for themselves

They’re all deep in AI, but every one of them has drawn a line where they’ve decided: this stays with me.

Ugnius’s line is clear. He won’t delegate the thinking about why something is broken.

Case in point: a recurring stability issue had been hitting his systems every few weeks. He could have patched it each time with AI.

But Ugnius isn’t the patching type. He dug through the configuration, the logs, what the application was actually doing, and eventually traced the behavior back far enough to realize the system was storing a lot of data it didn’t need.

AI can find a fix. It can’t decide whether the fix is worth finding, or whether the real problem is somewhere else entirely.

Ugnius can because he’s across multiple projects at Hostinger, not just his own. He sees how systems connect, and how a change in one place ripples through others.

Eylül has a similar instinct. She won’t let AI anywhere near architecture planning for new initiatives.

“I’ll use AI to help build individual pieces, but I don’t trust it to design and plan the whole thing for me.”

The overall design stays hers. She’s also the one who, despite having automated most of her workflow, still makes a point of talking directly to customers.

“There’s something really valuable about understanding who’s actually using what you build every day,” she says.

When everything else runs through agents, she’s deliberately keeping one channel human to stay connected to the bigger picture.

Karolis still evaluates YouTube channels manually. AI gives generic insights and sometimes misses the nuance.

His creator partners are developers, AI automation enthusiasts, ecommerce experts creating detailed tutorials and deep dives. Knowing what makes a good partner requires product knowledge, technical understanding, and a trained eye no model has matched yet.

“This will improve over time,” he says, “but for now, a manual eye check still works better, as we can validate the channel based on our own knowledge and experience.”

And Gabrielė? She keeps AI out of one specific place: team brainstorms.

“I still really value that human interaction and those random ideas that come from people bouncing thoughts off each other,” she says. “Of course, AI can join after the meeting to recap everything and help us organize the chaos. But during that creative part, I’m happy to let the humans do their thing.”

Too deep to go back now

For all the friction, they’re in too deep to go back now. And they know it.

Eylül said she’d be “devastated” if someone pulled the plug on AI. Gabrielė said going back to doing everything manually would “feel almost illegal.”

They like how they work now. The speed, the range, the fact that they can build something in an afternoon that would have taken a week.

They’ve found a way to do their best work, and it shows.

If that sounds like your kind of chaos, we’re hiring. There’s room at the bar. Gabrielė’s agent says there isn’t. There is.