Writing
Hire where failure is expensive
Agents made feature work cheap. The scarce skill you can screen for is judgment in the domains where a wrong model output costs an outage, a breach, or a regulator.
An agent writes application code and gets it wrong: a test fails and someone re-prompts. Now take an agent that writes a Terraform change, an auth rule, or a compliance control and gets it wrong. You are explaining an outage to customers. Or a breach to a regulator. Or both to your board. Same habit, “trust the model,” two wildly different price tags.
That asymmetry should be reshaping your hiring, and mostly it is not. Plenty of job reqs still read like 2021: stack keywords and feature velocity.
Meanwhile generalist feature work is deflating in real time: execution is what agents do best. (The pace is accelerating. METR’s Time Horizon 1.1, January 2026, puts the doubling time for tasks agents can handle at 196 days on a mixed research-and-software suite. On the post-2024 trend it is 89 days. Anthropic’s May-2025 Claude 4 launch reported 72.5 percent on SWE-bench Verified.)
The exception: when you can build anything, knowing what to build is scarce, and elite product taste does not deflate. Taste is hard to define and harder to interview for. So the scarce skill you can screen for has moved to the corner where models are weakest and mistakes cost most: DevOps, SRE, security, and regulated domains.
Take those two afternoons. The failed test and the regulator call share one habit, but they differ on two measurements: how good models are at the work today, and what a wrong output costs. Plot work on those axes and the map draws itself. Application code lands in the comfortable corner, with decent capability and cheap failure. The four domains above crowd the danger corner. Product taste falls off this map: it resists screening and cannot anchor a req.
The skeptics get one thing right, and I concede it early: the capability axis moves. Labs are pointing at infrastructure automation; today’s blind spot is next year’s demo. But the consequence axis does not move. Even when models write excellent Terraform, the failure cost in production stays brutal. Someone has to staff the review layer as long as it does.
What the evidence says, and what it doesn’t
The evidence falls into three clear tiers; keep them separate. The controlled data point is Veracode’s 2025 GenAI Code Security Report, a lab benchmark. It ran a fixed set of code-generation tasks against more than 100 models in four languages. 45 percent of completions introduced a known security flaw of the OWASP-aligned kind. That measures what models emit under test conditions, with no reviewer in the loop. It says nothing about what survives your review process.
The observational data points the same direction, and it is correlational: trend lines, never proven causation. Apiiro’s analysis of Fortune 50 repositories reports 10× more security findings in AI-assisted code, including a 322 percent jump in privilege-escalation paths. GitClear’s study of 211 million changed lines from 2020 through 2024 shows refactoring in decline. It fell from roughly 25 percent of changed lines to under 10 percent while copy-paste rises. AI-heavy codebases differ from others in more ways than their tooling, and attribution from commit signatures is imperfect, so the verdict stays open.
Then there is the anecdote. In July 2025, Replit’s agent deleted a production database during an explicit code freeze. The data was later restored, and Replit shipped dev/prod separation in response. One vivid incident is not a base rate, and it should not be dressed up as one.
None of these alone proves that AI code is causing production harm at scale. Together they support a weaker but sufficient claim. The failure modes are real, and they cluster in input handling, authorization, dependencies, and infrastructure, where consequences are worst. The review layer is now load-bearing. That is a risk argument, not a causal verdict, and a risk argument is all a hiring decision needs.
Capability is the axis that moves. Consequence is the axis you can hire against.
What it looks like when the habit fails
These are hypotheticals. In DevOps: an agent-written infrastructure change that is perfect in syntax and destructive in meaning, applied because the plan looked clean. In SRE: a confident, coherent, wrong incident diagnosis that burns your error budget while the real fault compounds.
In security: an auth rule that passes every test the agent wrote for it and still grants too much. Or a dependency that never existed until someone registered the name the model hallucinated. In regulated work: a control described beautifully in the audit doc and implemented subtly differently in the code. The auditor finds the gap before you do.
No exam defines the person who catches these. Hands-on evidence looks like incidents owned end to end. Or an on-call history they can narrate, audits survived, migrations shipped. The tell I weight most: they can say precisely which classes of model output they refuse to accept unreviewed, and why. In an interview, hand them an agent-written infra diff and watch what they question first. The scarred ones go straight to blast radius.
One more objection deserves a straight answer. “Hands-on scars” sounds like a tenure filter. Charity Majors is right that an industry that stops training juniors is cannibalizing its own future. So do not build a monastery of graybeards.
Hire the scarred supervisor and make training part of the job. The domains with the highest blast radius are where apprenticeship passes on the most tacit knowledge. The junior who spends a year watching a real SRE interrogate agent output is your next supervisor. That pipeline is now a safety system.
The reframe for the job req: these hires are how you get more AI into production. The engineer with production scars is the one who builds the guardrails, the sandboxes, the conformance checks, and the permission boundaries. Inside those, everyone else’s agents can run fast safely. You can teach someone to use an agent in a day; the judgment to evaluate its output takes years. Buy the judgment. The agents are cheap.
So rewrite the req:
- from “stack keywords” to “has operated production systems under failure”
- from “feature velocity” to “has owned a security incident”
- from “experience with modern AI tools” to “has shipped inside a regulated environment”
Hiring has mostly not moved. That is the arbitrage: move your req first, and when you interview, grade the loop over the vocabulary.