Rootly's Sylvain Kalache warns AI could dull on-call skills

Sylvain Kalache (@sylvainkalache), who built a self-healing infrastructure prototype at LinkedIn in 2012, is warning that today's AI incident tools could leave engineers unprepared for the failures those tools cannot resolve.

Kalache, now AI Labs lead and developer relations executive at Rootly, argued in a September 2026 essay that automating routine investigations removes the practice through which responders learn how production systems behave under stress. His prediction is specific: automation will lower average mean time to resolution, while complex incidents will take longer as engineers lose familiarity with the systems they operate.

That remains a hypothesis, rather than a measured industry trend. It is also an unusually direct warning from an executive at Rootly, an incident-response software company that markets AI SRE tools. Rootly benefits when customers automate more incident work. Kalache's argument is that the same customers will need a deliberate way to replace the experience automation removes.

The easy incidents were training

Routine outages are inconvenient, but they give engineers relatively forgiving opportunities to inspect telemetry, test hypotheses and learn which dependencies fail together. An agent that handles those cases at 2 a.m. saves sleep and engineering time. It also takes a repetition away from the human responder.

Kalache calls the resulting gap "comprehension debt": the distance between how a production system works and how well the people responsible for it understand it. When an unfamiliar, high-severity incident exceeds an agent's capabilities, the remaining human task is harder and the responder has had fewer chances to practice.

The argument follows the automation paradox described by human-factors researcher Lisanne Bainbridge in her 1983 paper, "Ironies of Automation". Automation leaves operators responsible for abnormal situations while reducing their exposure to the routine work that builds expertise. Bainbridge concluded that operators in automated environments require deeper skills and continued hands-on training.

Kalache's own route to that conclusion runs through education. After LinkedIn, he co-founded Holberton School, which taught software engineering through projects and peer learning instead of lectures. In his essay, Kalache recounts adding broken-infrastructure exercises after Dropbox said some Holberton graduates lacked troubleshooting experience. Students had to diagnose and repair the systems themselves.

That experience now shapes his view of AI-assisted operations. An agent can show its evidence and explain the path to a diagnosis, but watching the work does not reproduce the pressure of leading an incident with incomplete information and several people waiting for a decision.

Rootly applies the thesis through simulation

In a September 2026 essay, Kalache said Rootly had partnered with Uptime Labs, an incident-simulation and responder-training company, to apply that idea through realistic incident simulations. Engineers take the incident commander role during a simulated e-commerce outage, use observability tools and coordinate with LLM-powered stakeholders in Slack.

The simulated executives, customer-support staff and other colleagues require participants to practice communication, escalation and coordination. Kalache presents the exercises as a way to rehearse forming a diagnosis from incomplete information while keeping a response organized under pressure.

Rootly and Uptime Labs have not published independent results showing that simulation offsets the skill loss Kalache expects from automation. Rootly says its Academy program is available to Rootly customers, with existing customers directed to their account teams for access.

The metric that would test his thesis is resolution time for complex incidents, segmented by how much routine response work a team has delegated to agents. Average MTTR can hide that distinction. A large improvement across common incidents could coexist with slower handling of the rare failures that carry the greatest operational risk.

Rootly is selling both sides of the automation cycle

Rootly co-founders JJ Tang (@jjrichardtang) and Quentin Rousseau (@quentinrousseau) met at Instacart, where Tang worked in product and Rousseau was an early site reliability engineer. Rootly says they started the company after watching outages disrupt millions of dollars in orders while responders relied on fragmented internal tooling.

Rootly initially centered incident coordination inside Slack, according to a Slack developer account of Rootly's early product. Rootly has since expanded into on-call scheduling, status pages, retrospectives and automated investigation.

That direction mirrors the market. On September 2, Traversal made its Incident Workers generally available, describing agents that join incident channels and investigate without waiting for a human prompt. Incident.io markets an AI investigation product while emphasizing that its broader platform remains designed for human responders.

The Uptime Labs partnership gives Tang and Rousseau another offering as those agents assume more work. Rootly can automate common incident tasks and provide simulations for less familiar failures. The pairing serves an obvious commercial purpose, but it also addresses a credible operational problem that automation vendors have little incentive to emphasize: a lower page count does not prove that the humans still know what to do when the agent stalls.

Kalache is offering a constructive answer rather than an argument for keeping engineers awake to preserve their skills. Teams can rehearse rare failures without sacrificing production reliability, exposing customers to outages or turning routine on-call work into an informal apprenticeship.

Software organizations already run game days, tabletop exercises and chaos engineering programs. Rootly and Uptime Labs are packaging those practices into repeatable exercises with simulated people as well as simulated infrastructure. The harder task will be proving that performance inside a drill predicts performance during a real, ambiguous incident.

Kalache's warning lands because he has spent time on both sides of the problem: building systems that recover themselves and building a school around learning by doing. AI can remove a large share of operational toil. His bet is that engineering teams will have to spend some of the time they save practicing for the moment the automation reaches its limit.