Currently in my car I have the “check engine” light being on, but it’s ok, because I know what is the problem, my mechanic told me that it’s <something>, and that even though it’s not ideal, it can wait while he finds the part to repair it (apparently it’s not so easy to find).

There is something I don’t like about this though: if there is a NEW problem appearing, I won’t know about it, because this check engine light has only one state, and now it’s being used.

On a computer of course the situation is quite different. You have thousands of sources of notifications, warnings and potential errors, and you need ways to manage them, in order to avoid drowning in noise. One source of noise I find particularly annoying and stressful is the notification that you need to update something. I find it difficult because in many cases, you are torn: on one side you are attracted by the novelty and the possibility of the newer version of something being better, but on the other, there is always the real possibility that upgrading a component will introduce a new problem (in software development this is so common and problematic that it has its own name: a regression). I don’t know if it is because I’m OCD or whatever, but I often find that hard to manage, and with time, I have been moving quite clearly on the conservative side of things: for instance, there is no way you will find me eagerly upgrading to the new version of MacOS (Tahoe something I think), even though I know in advance that the upgrade warnings will annoy me. I will certainly wait for a couple of .x versions to pass. Some colleagues of mine were very early MacOS upgraders, and I was horrified by it. Why would you take the chance to render such a complex work tool useless because you are too eager to try a shiny new UI feature? By the way, I have a similar “subtly stressful” relationship with unread notifications, but that is another story.

While doing software development, there is a deeper peril related to this, for which I had two painful experiences recently, while using Claude and Codex to code. The first one happened when I asked Claude to give me some code for the build recipe of a Docker container:

ARG LO_URL="https://downloadarchive.documentfoundation.org/libreoffice/old/6.4.7.2/deb/x86_64/LibreOffice_6.4.7.2_Linux_x86-64_deb.tar.gz"
WORKDIR /tmp/lo
RUN wget -q "$LO_URL" -O lo.tar.gz \
 && tar -xzf lo.tar.gz \
 && cd LibreOffice_6.4.7.2_Linux_x86-64_deb/DEBS \
 && dpkg -i *.deb || true \
 && apt-get update && apt-get -f install -y

Can you spot the problem? I will make it easier for you, by showing you a second instance of the same problem, this time offered by Codex, while helping me code something in Python:

def api_request(
    method: str, endpoint: str, data: Optional[Dict[Any, Any]] = None
) -> Optional[requests.Response]:
    """Make authenticated API request"""
    try:
        url = f"{API_BASE_URL}{endpoint}"
        headers = get_headers()

        if method.upper() == "GET":
            response = requests.get(url, headers=headers)
        elif method.upper() == "POST":
            response = requests.post(url, json=data, headers=headers)
        elif method.upper() == "PUT":
            response = requests.put(url, json=data, headers=headers)
        elif method.upper() == "DELETE":
            response = requests.delete(url, headers=headers)
        else:
            return None
    except Exception:
        return None

Can you see it now? Yes of course, in both cases, the AI decided that it was a good thing (tm) to send any errors in a black hole (the first time with the true command, and the second time with the dreaded except Exception Python pattern), and both times, I lost a lot of time, wandering completely in the dark, because there was no “stop the world, an error occurred!” safety net. What is notable in those two cases is that it’s quite hard to come with any good reason for these two draconian error silencing mechanisms, in the first place. In both cases, if there is an error at this particular stage, it should be an immediate showstopper, and the underlying problem should be investigated in priority. Silencing these errors is a very risky strategy, and one thing is certain: it shouldn’t be the default solution that is being offered by an AI coding agent, whatever the context. Please AI coding agents, you should learn to NOT do that!