From the data centre floor to AI agents

My first job in tech did not involve much code. It involved a cold room in Jakarta, a trolley of servers, and a lot of cable.

I racked machines. I crimped UTP. I ran power cabling and labelled it, and then labelled it again because someone had moved it. On night shifts I sat in front of a KVM switch, flicking between consoles, watching for anything that looked wrong. When a client called at 2am, it was usually me who answered.

It was not glamorous. I loved it anyway.

The sysadmins

What made that place special was the people. A few of the senior sysadmins were, honestly, legends. One of them could walk past a rack and tell you which disk was about to fail. Another could read a log file the way most people read a newspaper, skimming, then stopping on one line and saying “there it is.”

They never seemed to rush. They also never guessed. If something broke, they wanted to know the physical path first. Which cable, which port, which power feed, which switch. Only then would they look at software.

I did not realise it at the time, but they were teaching me a way of thinking. Every system is a stack of real things. If you understand the layers, you can reason about failure. If you do not, you are just hoping.

Climbing the stack

After the data centre I moved up the stack. I wrote Java for banks, integrated with core banking, worked on switching and card networks, and later led architecture for health tech and telco platforms. Each step took me further from the hardware.

But I never quite forgot it. When a service was slow, part of my brain still asked whether it was the disk, the network, or the code. When a design review got abstract, I still pictured the boxes in the rack and the cables between them. That habit has saved me more than once.

Then the agents arrived

These days I work with AI agents every single day. They write code, draft documents, summarise logs, and help me write Linux and DevOps guides.

The first few weeks were exciting and slightly alarming. Agents are fast. They are also confident in a way that a junior engineer rarely is. They will tell you a config is correct when it is not. They will invent a flag that does not exist. They will fix the symptom and quietly move the problem somewhere else.

Here is the funny part. The skill I needed most was not prompting. It was the old night shift skill. Know the layers. Check the physical path. Do not trust a green light just because it is green.

Directing, not just asking

When I give an agent a task now, I treat it a bit like briefing a new colleague on the data centre floor. I explain the environment. I explain what good looks like. I explain what must not be touched. Then I check the work, layer by layer.

If an agent proposes a change to a network setup, I can tell whether it makes sense because I have physically plugged those things in. If it writes a database migration, I know what happens when a disk fills up halfway through. That foundation lets me evaluate output quickly instead of reading every line with suspicion.

It also lets me do something more valuable. Once I see an agent do a task well a few times, I can turn it into a repeatable use case. A clear input, a clear check, a clear owner. That is the difference between playing with AI and actually using it at work.

What the legends would say

I sometimes wonder what those sysadmins would think of all this. I suspect they would be curious rather than impressed. They would ask the agent where the data came from. They would want to see the logs. They would probably catch it making something up within five minutes, and then keep using it anyway, because it saves time.

That feels like the right attitude. Agents are tools, very good ones. They do not replace understanding. They reward it.

The thread

If there is one thread running through my 19 years, it is this. The technology keeps changing, but the discipline does not. Understand what is underneath. Be honest about what you do not know. Check before you trust. Leave things more reliable than you found them.

I learned that standing next to a humming rack at 3am. It turns out it works just as well sitting next to an AI agent at 3pm. The coffee is better now, at least.