The angle is that almost every real bug was in the plumbing or the config rather than the AI, and several looked exactly like success: the fabricated proof points that came from your own config file, the Blur hook that was true and sourced and useless, four bugs living between "a job is queued" and "a job runs", twenty green ticks with zero posters, and Cloud Trace I'd claimed rather than built.
That's an honest post rather than a marketing one, which is why it's worth reading.
source & further reading
dev.to — original article
The safety layer that had never run
Using an embedded feedback loop to develop with AI agents
I asked one model to review another, and it found a bias I would not have