Essay says AI coding has made developers architects without telling them

Essay says AI coding has made developers architects without telling them

The essay starts with a bug. The author was debugging a performance problem in their company's Google Forms addon and found that a colleague had vibe coded the logging part so he could monitor new installs every week. The author grants the goal is sound, since watching new customers is the most important part of running a business. The trouble was the method. The colleague was logging the user's email, and Google has severely restricted how much user detail an addon can get since last December. So the AI had written a series of try/catch statements: first try to get the email if the user consented to share it during installation; if that fails, try the user ID; if that fails, log an error saying someone installed but the email and user ID are unknown.

Each of those is one line of code, but internally each makes multiple API calls to Google servers. Because logging was the first thing to run after installation, the welcome screen took 15 seconds to load.

The author then asks what the right code would be. A senior developer or architect who knows how many API calls sit behind each method would have logged only the install timestamp and told marketing they cannot send a follow-up email based on installation alone. Or, if marketing insists on the email at install, show the welcome screen first and log the email or user ID asynchronously from the browser while the user looks at it. That route has its own catch: calling your server from the welcome screen might require the server to bypass security and accept an unauthenticated payload. In short, the architect's job is to push back on requirements that can slow the app, or to find a workaround, by making things async or accepting a reasonable hack.

Why AI will not do this, in the author's telling: it is great at writing code, but it does not know whether you are launching an app or have millions of users, and it needs that to choose between logging a timestamp and logging an email. It is also sycophantic and executes your requirements exactly without pushing back, which gives you a slow app that takes 15 seconds to load a static welcome screen. If you specify requirements rather than architecture, the AI decides the architecture for you. The author adds that it is so paranoid about security that it will never write code to accept an unauthenticated payload, and even if you write one yourself it will rewrite it into an authenticated one, hurting performance or forcing users to sign in again.

The author says AI coding tools are tuned to implement functional requirements plus two non-functional ones, security and backward compatibility, so as not to break your app. That invariably produces architecture that trades performance for security, and multiple versions of the datamodel with redundant tables, columns and code.

The prescription: you define the architecture, design the datamodel first, and ask AI to code within that constraint. Whenever a new requirement would change the architecture or design, the coding tool should have to ask before changing it. You also tell it when backward compatibility is needed, so keeping several versions of logic or tables is fine, and when to delete the old logic and tables and keep a single version for maintainability.

The author's broader claim is that developers who were great at coding have been moved into an architect's job. They see the work they love, coding, now done entirely by AI, and nobody has told them about the new job: pushing back on requirements, designing the datamodel, constraining AI to their architecture and reviewing that the code falls in line. The author calls this the reason for the confusion and hate among developers about vibe coding.

The closing is reassuring and blunt. In this case the right fix is simple: log the timestamp, show the welcome screen first and tell marketing what they cannot get on install. The coding job is going away, the author says, but that does not mean the developer's job is going away. Non-developers who do not understand architecture and design will end up with bad performance and unmaintainable database tables. So developers should stop missing the coding and learn to review code the way an architect reviews a developer's.

Key facts

  • AI-written install logging in a Google Forms addon made a chain of Google API calls (email, then user ID, then an error log) and delayed the welcome screen by 15 seconds.
  • The author's fix: log only the install timestamp, show the welcome screen first, and tell marketing what they cannot get on install. Alternatively, log email or user ID asynchronously from the browser.
  • The author says AI is sycophantic, does not know whether an app has few or millions of users, and decides the architecture for you if you specify only requirements.
  • AI tools are tuned for functional requirements plus security and backward compatibility, which the author says leads to slow code and datamodels with redundant tables, columns and code.
  • The advice: design the architecture and datamodel first, have AI code within that constraint, and tell it when to keep or delete old versions.

Why it matters

The essay offers a concrete answer to why many developers dislike vibe coding. The author's argument is that the job quietly changed from writing code to deciding what should be built and how, and that nobody announced it. The 15-second example makes the point tangible: the AI did exactly what the requirement said (capture the user's email) and the result was a slow, static welcome screen. The author's position is that coding is going away as a job but the developer's job is not.

Who it affects

Mainly working developers who used to spend their time writing code and now review or direct AI output. The author also names the marketing team as a requester, whose requests (such as an email on install) carry hidden performance costs that someone has to push back on. Non-developers building with AI are included: the author says those who do not understand architecture and design will end up with bad performance and unmaintainable database tables.

How to use it

The author's advice is a working method. Define the architecture and design the datamodel first, then ask AI to code within that constraint. Make the tool ask before it changes architecture or design when a new requirement demands it. Tell it explicitly when backward compatibility is needed, so keeping several versions of logic or tables is acceptable, and when to delete old logic and tables and keep a single version for maintainability. Review the result the way an architect reviews a developer's code. When a requirement is costly, push back on it, or make the work async, or accept a reasonable hack.

How solid is it

This is a first-person opinion essay built on one anecdote. The 15-second figure is a single observation from the author's own addon; no measurement method or after-fix timing is given, and the post does not say whether the fix has been deployed. The claims about AI behaviour, including sycophancy and security paranoia, are the author's assertions with no data, survey or study behind them. The post names no specific AI coding product and does not say which tool wrote the logging code. The author's name and the colleague's name are not given.

Risks and caveats

Generalising from one addon and one unnamed AI tool is a stretch; other tools may behave differently. The author's own example has a trade-off: the asynchronous route may require the server to accept an unauthenticated payload, which is the security compromise the author says AI will refuse to write. The framing also puts the blame on missing architectural judgement, not on the colleague's goal, which the author calls absolutely right. No timeline is given for when the coding job will go away.

“You got promoted. Nobody told you, so I am telling you.”

— From the essay on manidoraisamy.com