On AI, building things, and sharing them

I've been less active online lately, including writing less, due to a combination of things. But I felt this could be one of those brain-dump posts where I just loosely link some thoughts as I write them.

At the time of writing this post, last weekend we co-located the final big piece of our client applications monorepo, the web & desktop clients. Millions of lines of Android, C++, iOS, and TypeScript code, finally allowing for quick cross-platform changes via Bazel in a single repository.

The celebration for some of us was attending the BazelCon 2026 in Amsterdam. Great event, and even greater time with my colleagues. And my dog loves the Aspect Build's cute mascot plush that I obtained from their claw machine.

Lately, lots of things seem to be happening at the same time, up to the point that some days I finish work too exhausted to either code or direct some AI agent to do anything regarding my personal experiments. I don't think that coding was ever "the problem" at most companies ¹, but now the amount of AI-generated output really requires some adaptation, as we can't absorb all of it as we did before. I'd bet that most companies are now facing a verification problem, not a code generation one.

Work has been as interesting as it has been intense at times, but that is only half of why I've been quieter online. The other half is how I spend my remaining spare development time.

It feels both surprising and alarming that back in February I was questioning if we could automate AI usage (as in, anything beyond a list of prepared tasks), and a little over half a year later, I am already managing small groups of agents in agentic loops where I can be as hands-on or hands-off as I want, at work and also for personal projects. AI agents still fail at times, they still produce some slop without steering and linting and guidance, and I still like to review the code.

But I am also already searching for better ways to verify their output, because good AI guidance and AI skills alone are not enough. They produce too much, too quickly. The automation part is more than feasible now, but the quality part still remains. I could perfectly be in the drawing depicting the fourth phase from my previous blog post, if I were not trying to multi-task:

What feels strange, maybe even a bit sad, is that lately I am feeling less incentivized to do open experiments, at least until I feel they are ready. And even then I need to decide if I will simply publish the end artifacts. When it is so easy to clone something from the source code, or even without it, what's the point in publishing? AI makes it possible to generate copies so quickly that small companies could be in danger of being replicated ². Are we going to stop sharing source code?

Closely related, I have not made any of them public, but I now have a growing list of personal apps and tools built via agentic development ³, created to mimic and replace existing alternatives, although with fewer features and less polish. At times it is a Dockerized website, other times a pyinstaller Windows executable, a cross-compiled Go GUI-based application, and soon an iOS app or two. But in all cases, to me it feels like the merit is blurry, as I did not really build them: I came up with an idea, defined it, prepared the tasks, and let my little robots do the work to then verify and indicate corrections. I am "building something", and yet I don't know how much of it is truly mine versus how much I simply delegated. If I don't feel particularly accomplished, why would I want to share it?

But it could also be a matter of changing the mindset. There are at least two experiments that I currently enjoy working on, and might end up being small cool "things" to share. One is forcing me to research to understand whether the output is as good as it looks, as it does fulfill my specification and is working nicely, but half of the code is from an area I am not familiar with (Machine Learning). The other is around game development, and I'm having a blast evolving something almost from scratch to who knows how far, finally experimenting with an idea I had for a while, now with the extra twist of having AI agents do a lot of the coding (although here I'm reviewing every single line). In both projects, the agents aren't merely executing a backlog. Their output gives me new things to investigate and decisions to make, and I often discuss approaches and trade-offs before and after the implementation, so the process feels much closer to working with a small development team.

Maybe a way to find more joy in working with AI could be (again) more human collaboration and/or mentoring others? If only we weren't all so busy working on so many things that even pairing feels unrealistic...