Humans and agents

This is my first post about AI, actually, my first post since AI transformed the software engineering sector.

Personally it isn’t even that bad for me: while big tech cuts employees by the thousands to transfer human staff expenses into the AI infra race, the company I work for hasn’t gone through the same process, for now. That said, it really does feel like the software market steered in a direction from which it will never turn back.

The bright side, for the engineers that still have a job, is that showcasing AI fluency has become an important factor in performance reviews, and you get unlimited tokens for the trial period.

What really changed the rate of adoption for me has been tools like Claude Code. I haven’t used in-IDE agents for a while; where CLI agents really shine is extreme automation. We can now easily develop without looking at code.

With caveats. Vibecoding is a buzzword, maybe it’s just me, but when I hear the term I envision somebody playing a slot machine until an app works. First, this isn’t engineering. Second, at the current state of things, that process might produce a static website, not more.

On the other hand, what does work is a certain set of techniques where you push the model towards state-of-the-art software.

When the relationship gets sick

The narrative of the people selling AI is “code is solved”. AI assistants are a bulky presence, and when you have a coding assistant like CC it’s hard to slow down, the adrenaline rush of producing an enormous amount of code is strong.

Going fast works when you have a strong knowledge of the context: you give accurate instructions to your favorite spec-driven dev framework, build up an accurate plan, and immediately spot if something is off track. Iterate, deliver.

But what happens when teams don’t have strong context? Usually bad things. Developers race to a POC, burn down the Jira board, and companies then complain that the AI spend isn’t producing a business return.

Can agents fix coordination too?

The obvious takeaway is that going faster is a challenge to communication. But maybe this is a challenge we will, in the end, be able to manage with AI.

Imagine we could automate — and therefore offload, the burden of prioritizing the things with the most valuable outcome. A system that ranks work by expected impact rather than by who shouts loudest.

I’ve been thinking about this for a while, but maybe in this very moment I just clarified what I want to build next. Tomorrow will be another day, but the spark I leave in these few paragraphs might live on.

I tend to be optimistic. I love conclusions with an uplifting note.