AI isn't replacing SREs, it's deskilling them

Original Article ↗ Hacker News Discussion ↗

AI-powered tools are increasingly handling routine operational tasks for Site Reliability Engineers (SREs), from log correlation and incident triage to automated remediation, raising concerns that humans are losing the hands-on experience needed to debug novel, complex failures. Many compare this to pilots over-relying on autopilot: automation boosts day-to-day safety and productivity, but can erode the skills required when systems behave unexpectedly or supporting AI services fail. Others counter that this “deskilling” is just a shift in emphasis — away from low-level toil and toward higher-level system design, incident simulation, and directing AI agents — but worry that without deliberate practice and training, future teams may lack enough deep expertise to safely run critical infrastructure.

AI use in SRE and perceived deskilling

  • Many see clear deskilling: people reach for Claude/ChatGPT first, copy‑paste between agents during incidents, and lose the habit of “using their brain.”
  • Others argue it’s just shifting the locus of skill: from writing syntax to writing specs, tests, and designing systems for observability and resilience.
  • Some note this “tiering” effect: AI handles routine/Tier‑1 and Tier‑2 incidents; humans only see novel Tier‑3 failures, which may erode intuition built from handling simpler cases.

Effects on skills, learning, and juniors

  • Several worry about long‑term atrophy: if AI is used from the start, future engineers may never learn to read, debug, or deeply understand code and infra.
  • Concern that the pipeline to Senior/Staff levels breaks if juniors mostly prompt and review, rather than build and debug from first principles.
  • Others say skills can be reacquired, and deliberate “practice time” (coding, puzzles, simulated incidents) can mitigate deskilling.

Debate over what SRE/DevOps should be

  • Strong disagreement on whether SRE/DevOps are:
    • glorified sysadmins / tool jockeys tied to Kubernetes, Terraform, YAML, or
    • software engineers focused on reliability, scalability, and cost.
  • Some insist “real DevOps” means devs owning on‑call for what they build; others argue ops is a distinct, valid specialization.
  • Large‑org reality: central SRE/infra teams persist to standardize platforms, enforce reliability, and handle cross‑cutting complexity.

Experiences with AI in operations

  • Multiple practitioners report frontier models are already excellent at:
    • log/metric/trace correlation, incident triage, CI failure analysis, and generating dashboards/runbooks.
    • CLI‑driven tasks (e.g., AWS) and end‑to‑end infra changes following checklists.
  • Others report agents misinterpreting alerts, chasing wrong layers, rewriting too much code, or making debugging harder when their earlier changes were subtly wrong.
  • Consensus: AI is powerful but only as good as context, tooling, and human supervision; incompetent people remain incompetent with better tools.

Analogies and historical parallels

  • Aviation is repeatedly cited: autopilot improves safety but requires mandated manual practice and training; people suggest “incident simulators” / chaos exercises for SRE.
  • Counter‑arguments: software lacks aviation‑style regulation and incentives, so most orgs won’t invest in mandatory practice.
  • Other analogies: calculators and long division, CNC vs manual machining, Docker/Terraform already “deskilling” infra work, and broader automation of mundane skills over history.