Out of SyncWhat we expect from AI, and from each other

I taught myself to code in 2019 the classic way, with tutorials, documentation and a lot of trial and error and I've worked as a developer since 2020. At the end of 2022 ChatGPT came out, and that's where the whole AI thing started. When I was hired into my current job, that I started April this year (2026), it was partly because I had been an early adopter of AI tools and got pretty good with it. It was clear to me early on that this would become the tool of my trade, whether I like it or not. I use coding agents every day, and I see the benefits, but also the risks.

I want to write about some observations I made at work over the last years. One of them started long before AI showed up: you often don't know what the other side expects from you.

For me that's personal. I've had big problems with impostor syndrome, and I still do. I think many people have. Underneath that I also have an inferiority complex. So when I don't know how much my boss expects from me, I tend to assume it's more than I'm giving. I think it's the same the other way around, in a different way. A manager doesn't always know what to expect from the people they hire, especially when they don't know them well yet.

That's a human thing, and in my experience it doesn't really go away. Lately I have the feeling that AI makes it worse.

A third party at the table

AI moves the expectations, and it moves them a lot. Every week there's a new model, a new tool or a demo where someone builds an app in an afternoon. I have the feeling each of those shifts what people expect a little, and not for everyone at the same time or in the same direction.

On the other side I see developers who haven't really tried AI yet, or only parts of it, or who tried it with a weak model, got disappointing results and put it aside.

I think a lot of this comes from the promise the industry makes: that it just works. That's the advertisement. In my experience it doesn't. Using AI well is a skill, and it takes time to learn, also for developers. There are a lot of other variables, but I think this is an important one.

My own expectations shift too, with every new model and every new harness. Things change all the time, and what I expect of the tools and of myself changes with them.

Now I'm the manager

When I work on something where I know the concepts well and use a lot of AI, a lot happens very fast. Sometimes I notice that I'm losing my grip on the code. I have trouble understanding what got written, and I really hate that feeling, when the code leaves my hands.

It's like being a manager with very eager people under you, who work so fast and write so much that you lose track of what they're doing. That doesn't feel good, and it's bad for the code.

AI also makes it easy to do several things at once, and that feels fast. There's a study from METR where experienced developers were actually 19% slower with AI, but thought they had been 20% faster.1 No idea how much of that is true for me. But when I run several things in parallel, I can't give any of them my full focus. And even on days when I keep track and get more done than I could have before, I end the day completely drained.

And then I still feel like I'm too slow, because I'm not sure whether I'm meeting what my boss or my colleagues expect.

What we don't say

I think talking about it is the fix. What makes the talking harder is shame.

On Reddit or YouTube I see engineers complaining about pull requests that were fully generated and never checked, which someone else then has to deal with. I don't want to be that person. I still think understanding the code is worth it. Sure, someone without a technical background can probably ship something with AI. But I don't think anyone can really maintain it later.

At the same time, many developers say in public that they review everything their agent writes, and I'm not sure that's always true. In a survey from Sonar, almost all developers said they don't fully trust AI code, but only about half said they always check it before committing.2

I suspect it goes in other directions too. Someone might use more AI than they let on, or want to use more and not say it. So even when people talk openly, I think they sometimes leave a little out.

Agreeing in private

I did have this conversation, with my manager and with some colleagues. The people I talked to agreed. Still, I have the feeling that most of us handle it on our own, and that makes it worse.

When two people agree in a one-on-one, the rest of the team doesn't hear it. I think it helps to talk about it where everyone hears it at once, in teams and across the company: what we expect from the tools and from each other, and what is okay to admit.

For my own work I also write things down. Over time I've built up written conventions for how my code should look: naming, structure, patterns, what goes where. The agents work against those. The conventions don't change with anyone's mood or with the news of the week, and at least some of what I expect is written down. I don't check every line, and I don't accept everything blindly. I check where it matters, and I adjust the conventions when the output drifts.

That doesn't solve the problem for a team. It's the part I can do myself.