🔥 Trending on HN

If AI handles the easy incidents, who practices the hard ones?

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
SRE

A professional who helps keep software services reliable.

deskilling

Losing ability because a skill is practiced less.

human-in-the-loop

People stay actively involved while AI helps with the work.

What happened

On February 28, 2026, SigNoz (a company that builds tools for watching software systems) published an essay in its newsletter, Observability Real Talk. The essay examines AI and SREs (site reliability engineers). It does not mainly ask whether AI will replace SREs. It asks whether heavy AI use could make SREs less capable during rare crises. The essay imagines AI handling 95% of incident response. That figure is a scenario, not a measured result for every team. Read the original SigNoz essay.

The story received 286 points and 212 comments on Hacker News, a site where technology readers share links. Those figures measure attention. They do not prove the essay is correct.

The background: automation changes practice

The essay draws on Lisanne Bainbridge, a cognitive psychologist, and her 1983 paper, Ironies of Automation. It says automation removes routine work while humans remain crucial when machines fail. Yet humans get less practice. They face rare crises with fewer chances to build judgment.

For SRE work, the essay lists AI help with alert grouping, known playbooks, log patterns, anomaly flags, cause suggestions, and familiar fixes. It also describes a three-tier model: known failures can be automated; partly understood failures get AI suggestions and human checks; novel cascading failures stay human-led.

Why the concern matters

If AI summarizes logs, suggests causes, and maps dependencies, SREs may stop forming hypotheses or understanding systems deeply. Visible output can improve while practice fades. The essay calls this deskilling.

For junior SREs, it uses never-skilling. They may never get enough hands-on incidents to build intuition. Less practice makes people lean more on AI. More reliance can create even less practice. A new failure may not match the patterns AI has seen.

What the source supports

The essay cites a medical example. It says endoscopists who used AI for polyp detection later fell from 28% to 22% on unassisted detection. It also cites aviation research about heavy autopilot use. It says the FAA requires more manual flying time. These are examples presented by the essay. They are not direct measurements of SRE skill loss.

The proposed response is not abandoning AI. It is deliberate practice, including some cases people could automate. It favors human-in-the-loop work. In that model, people investigate with AI, rather than merely approving its answer. SigNoz says its own product direction follows this approach, with human control, guardrails, and gradual autonomy.

What remains unknown

This is one newsletter essay from a company that builds observability products. It does not establish how often SRE deskilling happens. The 95% scenario is not a benchmark. The cited studies do not show that every SRE will lose skill. We also do not know which practice rules work best for beginners or rare failures.

What to watch next

Teams may reveal the answer through their habits. Do they automate familiar incidents while reserving practice time? Do junior engineers read logs and form explanations themselves? During a new outage, does AI help people think, or replace the thinking? The important question is not whether teams use AI. It is whether they can use it while keeping human judgment ready for the cases that matter most.

💬 HN discussion: Does AI replace SRE work or erode its skills?

The comments do not reach one conclusion. AI may speed up incident response, but it may also encourage people to skip understanding how systems work. The incidents, numbers, and test counts below are commenter self-reports, not independently verified findings.

  • A strong counterargument is that AI may already be useful for SRE work. One commenter judged it fairly capable at correlating metrics, logs, and traces, performing autonomous root-cause analysis for partially understood failures, and assessing change risk. This was an opinion, not a comparative benchmark.
  • There was also a failure case. One commenter self-reported that a team used Claude first and spent two hours debugging a bug that should have been obvious. Even a capable tool can waste time when the first diagnosis is delegated to it.
  • One proposed safeguard is to write a detailed specification before prompting the agent, define the libraries and methods to use, and run tests afterward. The idea is to keep design and verification under human control.
  • The deskilling concern also had a counterexample. One commenter self-reported not programming from 2000 to 2015, then returning without much difficulty. That person was more worried about people who never learned programming before relying on an LLM.
  • Kubernetes was criticized as a high-level abstraction that can hide SRE and DevOps fundamentals and turn infrastructure into a black box. Another commenter countered that a well-designed platform, with logging, instrumentation, databases, deployment paths, and IaC, can remove accidental complexity.
  • Some commenters argued that making engineers on-call for the code they write gives them a reason to inspect AI-agent output and understand how it works.
  • As an example of testing complexity, one commenter self-reported locally simulating a storage system built on a minimal Linux ISO, with about 50,000 lines of Rust userspace code excluding about 10,000 lines of init system, about 2,500 unit tests, about 110 fault-injection tests, and about 60 block-store failure tests. This is one person's engineering example, not a general performance comparison.

initial digest at 212 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Can AI help SREs too much?

📰 Full story: If AI handles the easy incidents, who practices the hard ones?

AI may fix familiar problems faster, but people still need practice.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
SRE

A person who keeps computer services working.

deskilling

Losing skill because of too little practice.

💡 The gist

  • AI can handle many familiar computer problems.
  • People still face rare, complicated breakdowns.
  • Less practice might weaken their skills.

SigNoz (a company that builds software-watching tools) published the article on February 28, 2026. An SRE (site reliability engineer) keeps computer services working.

The article says AI can group alerts, find patterns in logs, and follow known repair steps. These jobs repeat often. AI can help people finish them quickly.

Rare failures are different. They may involve several systems. They may have no familiar solution. People must read clues, test ideas, and decide what to do. If AI always performs those steps, people may practice them less. The article calls this deskilling. It means losing strength in a skill through too little practice.

The article connects this concern to a 1983 paper about automation. Machines can handle normal work. Humans then handle unusual failures. However, unusual failures happen less often. People get fewer chances to learn them.

The article also discusses beginners. A new SRE needs repeated experience. If AI handles common incidents, that experience may disappear. A beginner might never build strong instincts. This is deeper than forgetting a skill. It means never developing it.

The Hacker News post received 286 points and 212 comments. These numbers show attention. They do not show that the claim is true.

The article cites studies from medicine and aviation. It says one medical example found unassisted detection falling from 28% to 22%. It says aviation rules require more manual flying. These examples support a warning. They do not prove every SRE will lose skill.

The article suggests keeping practice time. People can work with AI during exercises. They should investigate, not only approve suggestions. Skill grows through repeated decisions. Automation can make today's work look better. It may hide tomorrow's weakness. The original SigNoz article offers one company's view. It does not settle the whole debate.

💬 What commenters say about AI and SRE work

The comments contain both warnings and counterarguments: AI can help with SRE tasks, but careless use can weaken understanding. The experiences and numbers are self-reported, not the results of an independent study.

  • AI may be good at some SRE tasks, such as combining metrics, logs, and traces, finding causes, and judging the risk of a change. But one commenter self-reported that a team asked Claude first and spent two hours on an obvious bug.
  • A safer method is to write a specification first, choose the libraries and methods, and test the result afterward. Making engineers responsible for their own code during on-call duty may also encourage them to check what the AI produced.
  • Some people worry that coding skills will weaken. Another commenter self-reported stopping coding from 2000 to 2015 and returning fairly easily. That person was more worried about beginners who start with an LLM.
  • Kubernetes, a tool for managing many servers, can hide important details and become a black box. But a platform with good logs, databases, deployment paths, and IaC, meaning infrastructure described as code, can reduce unnecessary complexity.
  • One commenter self-reported testing a minimal-Linux storage system with about 50,000 lines of Rust, excluding about 10,000 lines of init code, about 2,500 unit tests, about 110 tests that deliberately cause failures, and about 60 storage-failure tests. This is one example, not a general comparison.

initial digest at 212 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

When a computer helper does the watching

📰 Full story: If AI handles the easy incidents, who practices the hard ones?

AI can help people watch computer systems.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
SRE

A person who keeps computer services working.

SigNoz

A company that makes tools for watching computer systems.

An SRE is a person who keeps computer services working. SigNoz is a company that makes computer-watching tools. Its article asks a simple question.

AI can spot familiar problems. AI can follow familiar fixing steps. That can save people's time.

But strange problems are different. People must think and try things. If AI always does the thinking, people practice less. Less practice can make skills weaker.

Hacker News is a site for technology stories. It gave the article 286 points and 212 comments. Many people noticed it. Those numbers do not prove the story is true.

The article says people should still practice. They can use AI while thinking together. The article is a warning from SigNoz. It does not prove every SRE loses skills.

💬 A simple story about AI and the people who run servers

These are comments from people, not a controlled study. The stories and numbers are what the commenters said about themselves.

  • Some people say AI is good at looking at many clues about a server problem. But one commenter said a team asked AI first and spent two hours on an easy-to-see bug.
  • People suggested making a plan first and testing the AI's answer. If someone is responsible when their own code breaks, they have a reason to check the answer carefully.
  • AI might make coding muscles weaker. But another person said they stopped coding from 2000 to 2015 and later came back fairly easily. They worried most about people who only learn with AI.
  • Kubernetes is a big server tool that can hide small details. A well-made tool can also make work easier. One commenter said they self-reported testing a Rust system of about 50,000 lines with about 2,500 normal tests, about 110 failure tests, and about 60 storage-failure tests. Those numbers describe one person's example.

initial digest at 212 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

Sources