HarpyWatch's writing on monitoring, uptime, and the things that break quietly.
Ask an agent for a login flow and you tend to get a ?next= parameter redirected to whatever it contains. It is the obvious implementation, it is an open redirect, and the remedy was documented years before the model was trained.
A catalogue of small, boring, real failures — a www that resolves nowhere, a redirect that drops to http, a response that is not actually compressed — each one cheap to see from outside.
An AI agent that ran certbot once did the work correctly and made no commitment about the renewal. The setup is in your repository; the recurring obligation it created is in nobody's calendar.
An outage costs what it costs. Being informed of it by the person it happened to costs something separate, it accrues, and it is charged against the next incident rather than this one.
Four agents changing one codebase means four diffs nobody has time to read end to end. The answer is not fewer agents: it is checking the deployed site, which costs the same however many changes arrived.
Generated markup is verbose, imports whole libraries for one function, and inlines assets because that works. Every test still passes. The cost lands on a phone on a slow connection, months later, as a number with no cause attached.
A strict Content-Security-Policy is usually the thing standing between an agent and the feature it was asked to add. Widening it is the path of least resistance, it is never mentioned in the summary, and nothing about the site looks different afterwards.
Adopting monitoring for a small site is five decisions, not a project. Here is what each one actually involves, and why finishing a rough version beats designing a perfect one.
A canonical tag pointing at staging, a noindex that shipped, a test API key in production. Each is invisible on a developer's machine and expensive live, and copying config between environments is the fastest way to make something work.
The instinct is to add monitoring once a site has traffic worth protecting. That gets it backwards: an empty site is the one failure mode nobody reports, because your first hundred visitors leave instead.
One agent edits the backend, another edits the frontend, a field quietly changes type, and every build stays green. The screen renders blank where a number belongs, and only something watching the deployed endpoint will tell you.
Every restructure moves URLs, and every move adds a redirect. Nobody ever removes one. The result is a chain a crawler stops following and a browser spends a second on before it sees a single byte of content.