AI is Responding to Incidents. Developers are Becoming Obsolete.

In a recent post, Sylvain Kalache, ex-LinkedIn SRE & Head of AI Lab @Rootly, describes an emerging concern: With AI taking over incident response (i.e. troubleshooting and fixing service incidents), engineers lose the “muscle memory” needed to stay fluent in the inner workings of their systems. Posted on September 4th, 2026. Original article here.

1The Aviation Analogy

Analogy with airplanes: the more you rely on automation (autopilot), the worse you get at manual intervention during crises. In the case of TransAsia Airways Flight 235, the plane crashed 117 seconds after the initial alert because the pilot had turned off the wrong engine. Unplanned failures in modern jet engines happen less frequently than once every 100,000 hours of flight. As a result, pilots rarely experience actual failures. To address this, regulators required pilots to take a simulator test every 6 months.

2Comprehension Debt

There is no such equivalent requirement in IT. As mean time to resolve (MTTR) drops for most incidents, it will seem like we’re making progress. However, we will be piling up a silent debt: when complex things break, and AI doesn’t know what to do, we will end up staring at systems that we don’t understand anymore from within. This is what I call “comprehension debt”. We can only get better at debugging by doing, and doing is moving to AI.

3What to Do

I believe that we need to require incident simulations for on-call engineers, just like in aviation. If we want to maintain our system fluency in an automated world, we need to stay fluent through practice. Without AI, we will be lost in chaos. With AI, we will be blind.