AI isn't replacing SREs, it's deskilling them
TL;DR Highlight
Letting AI handle 95% of incidents may dull the skills SREs need for the critical 5%.
Who Should Read
SRE or backend engineer adopting AI incident-response tools who worries about on-call skill erosion in the team
Core Mechanics
- More automation, weaker humans
- Bainbridge's 1983 paper Ironies of Automation described this paradox.
- Humans are only called in for rare high-stakes failures, but with less practice they handle them worse.
- PagerDuty's three-tier incident model
- Tier 1: known issues with known fixes are fully automated.
- Tier 2: AI recommends and a human validates.
- Tier 3: novel cascading failures stay human-led, with AI providing context.
- If humans only see Tier 3, where does intuition come from?
- Intuition comes from years of hands-on incident response and repetition.
- Offloading log reading, hypothesis forming and dependency mapping to AI stops that practice.
- Other fields show the same decline
- Medicine: endoscopists' unassisted polyp detection fell from 28% to 22% after a period of AI use.
- Aviation: heavy autopilot use degraded pilots' situational awareness and manual flying, so the FAA mandates more manual flying time.
- Deskilling hides behind productivity
- AI makes tasks feel easier and boosts visible performance, so people misjudge their own skills.
- Dashboards stay green until a novel incident exposes the atrophy.
- Never-skilling: skills never formed
- Deskilling is losing a skill you had; never-skilling is never developing it.
- Juniors lose chances to build foundational intuition because AI handles incident response.
- The author's proposed countermeasures
- Practice deliberate inefficiency: designate some AI-solvable incidents as human practice.
- Build human-in-the-loop where people engage with the problem, not just approve AI output.
Community Reactions
- Aviation is a fitting analogy
- Commenters said aviation has learned how to keep humans from erring as automation grows.
- One comment said the US Navy decided 50+ years ago that a machine may suggest but a human must carry out the action.
- Counterpoint: Terraform deskilled us too
- One commenter argued AI is just a new tool, as nobody asks SREs to configure infra by hand.
- They also claimed five nines needs automatic detection and remediation.
- 10x faster with AI, but atrophy feels real
- A startup developer said Claude with read-only access solved AWS issues about 10x faster and wrote reports.
- They admitted it sometimes BSes and said skill atrophy is real and worries them for job hunting.
- Tier 3 intuition only comes from Tier 3
- One commenter with complex production debugging experience said the key intuition is persistence and testing hypotheses, built only by facing Tier 3 issues.
- Another said at an org with most noisy pages automated, only novel faults remained, so intuition built more slowly.
- Practical coping suggestions
- One comment suggested a weekly coding hour, LeetCode questions, and having the agent quiz you before merging.
- Another said Claude needs a harness built from company context, and without your own knowledge it's a blunt tool.
How to Apply
- Solve one incident by hand first this week
- Today, pick a past incident and read the logs yourself before viewing any AI summary, then write down 3 hypotheses.
- Compare them with the AI's result and note the differences.
- Mark practice incidents in the on-call rotation
- Designate some Tier 1 incidents as human-handled practice cases.
- Have juniors write a hypothesis before seeing the AI's suggestion.
- Replace approve buttons with hypothesis input
- In Tier 2 workflows, make the AI only recommend, not allow one-click approval.
- Require the owner to enter a cause hypothesis and evidence before proceeding.
- Add an agent quiz before merging
- Before merging an AI-written change, have the agent ask the owner questions about it.
- Hold the merge if they can't answer.
Try It
# Paste into a Claude or ChatGPT chat (simplified version of the article's method: turns the 'humans engage with the problem' idea into a prompt)
You are an SRE on-call coach. While I work an incident, do not give me the answer; help me form hypotheses.
[Incident]
{{symptoms, alerts, recent deploys}}
[Rules]
1. First ask me one question so I choose which logs/metrics to check.
2. Do not state the cause until I write a hypothesis.
3. When I write one, give counter-evidence and how to verify it.
4. At the end, summarize what I missed in 3 lines.
Start.
Next step: compare against the answer AI gives when it solves the same incident directly, and note how close your hypothesis was.Terminology
Related Papers
Backdoor in Claude Code for VS Code extension on the Cursor Marketplace
A tampered Claude Code extension on Cursor Marketplace planted an SSH backdoor
Building an AI-native engineering org
Automating each shipping step raised merges per engineer about 5x with reverts under 0.5%
We ran AI agent swarms in Minecraft and Mindustry
We put 50 agents in Minecraft and they invented ration ledgers and bureaucracy.
The Agent SDK
Cursor released Agent SDK, letting you define agents in code and deploy them with one command.
Here be monsters: What do you still need to know when agents do the work?
Owning agent-written code depends on fundamentals, breadth, and diving deep just in time.
AI promised less work. I've never worked more
Running agents in parallel brought my GitHub activity back to full-time engineer levels, and I'm exhausted.
Show HN: Corgi, a helpful {agent, terminal} multiplexer
A tmux-like terminal tool that shows the status of all your AI agents at a glance
Ask HN: If the models are so good, why doesn't OpenAI just start 100 companies?
HN answers: code is easy, knowing what to build isn't, so they sell tokens instead