Former LinkedIn SRE warns AI incident handling erodes engineers' system knowledge

A former LinkedIn site reliability engineer has published an argument that is drawing attention on Hacker News: as AI tools take over routine incident response, engineers are losing the hands-on practice that prepares them for the failures automation cannot fix. In the post, Sylvain Kalache — who built a self-healing prototype at LinkedIn in 2012 and now works at incident management platform Rootly — frames the trend as a quiet skill-erosion problem the industry is not budgeting for.

What the new tools actually do

Kalache describes modern AI incident responders, commonly marketed as "AI SREs" (a term he says he dislikes), as end-to-end: they read alerts, hypothesize about causes, query telemetry, tie incidents to recent deployments and in some cases apply the fix themselves. He is candid about the appeal — a routine capacity issue handled overnight without waking anyone is hard to complain about. His concern is what those routine incidents used to provide: a relatively safe venue where responders built intuition for how their systems behave and how they fail. The better the tools get at routine cases, the less of that practice remains.

An irony identified in 1983

The dynamic is not new. Kalache points to human-factors researcher Lisanne Bainbridge and her 1983 paper "The Ironies of Automation," which argued that automation strips operators of chances to practice routine work while still holding them responsible for abnormal situations — and that operators therefore need to be more skilled and better trained than before, not less. Kalache's prediction follows directly: average MTTR for most incidents will fall as AI-assisted response spreads, while resolution times for complex, never-seen-before incidents will climb, because the humans who take over will be investigating systems they no longer know intimately.

The aviation parallel

He looks to commercial aviation for how to respond. Automation handles most of flying, but pilots remain accountable for the events it cannot manage: engine failures, unreliable instruments, rejected takeoffs, stalls. Such events are extraordinarily rare — modern turbine engines average fewer than one in-flight shutdown per 100,000 engine flight hours, so a pilot can complete an entire career without meeting one outside a simulator. When they do occur, there is no time to learn. Kalache cites TransAsia Airways Flight 235, where the right engine's propeller auto-feathered shortly after takeoff; the aircraft was designed to keep flying on its left engine, but the crew misidentified the problem and the plane stalled and crashed 117 seconds after the first warning. Aviation's answer is mandatory rehearsal: under US FAA rules, captains complete recurrent training or a proficiency check every six months, including scenarios such as an engine failure during takeoff.

Practice beats explanation

Software incidents rarely threaten lives, Kalache notes, but that is no reason to skip the craft. His proposed fix is simulation. Rootly has partnered with Uptime Labs on realistic incident drills in which an engineer takes the incident commander's seat during a simulated e-commerce outage, working observability tools while coordinating in Slack with stakeholders — a CEO, customer support — played by LLM-powered agents. He also considers using AI as a trainer, asking an agent to explain the steps it took and the evidence behind its diagnosis, but argues that explanation is not practice: watching Serena Williams does not teach you tennis. His conviction comes from more than five years running a software engineering school built on learning by doing — when Dropbox reported that its hires from the school were still weak at troubleshooting, he responded with projects that handed students broken infrastructure to diagnose and repair.

Comprehension debt

The post gives the risk a name: "comprehension debt," a growing gap between how a team's systems actually work and how well its responders understand them. His recommendations are concrete — regularly interact with the systems you watch over, deliberately handle unfamiliar failures, practice working under pressure, and rehearse the coordination and communication a SEV0 demands, through tabletop exercises and chaos engineering. Those techniques predate LLMs, he acknowledges, but they matter more now. Bainbridge reached the same conclusion decades ago: operators need regular hands-on control and simulation precisely so their skills do not decay.

Why it matters

AI incident response is being adopted quickly because its payoff is immediate and easy to measure: fewer pages, shorter MTTR, less toil. The cost Kalache describes is the mirror image — diffuse, delayed and invisible until the exact moment it matters, when automation hits a wall and an under-practiced human has to take the keyboard. Teams rolling out these tools should treat simulator time and hands-on drills as part of the automation's total cost, the way airlines treat recurrent training as part of flying. Otherwise the dashboards will show averages improving while the tail risk — the rare, severe, novel failure — quietly grows behind them.