Charity Majors: AI Demands More Engineering Discipline, Not Less

Charity Majors, co-founder and chief technology officer of the observability company Honeycomb, argues that AI-generated code does not reduce the need for engineering discipline — it sharply increases it. Her essay, published on June 15, answers readers of her earlier piece on AI enthusiasts and AI skeptics.

What changed in 2025

Majors dates the shift to last November, when the release of Opus 4.5 showed that AI could generate code roughly as good as that of the median software engineer for common patterns, and do it faster and more cheaply. Agentic harnesses, tool use, function calling and MCPs had been building through 2025 and crested into general-purpose usability at the end of the year, she notes, making the model release a tipping point rather than the cause.

Her central claim is economic: the economics of code production were turned upside down. Writing code went from hard, slow and expensive to effectively free and instant, and lines of code stopped being an asset to be curated and became disposable and regenerable.

Code as a cache, not an asset

Majors borrows the frame from Chad Fowler, who coined the term "immutable infrastructure" in 2013 and now writes about Phoenix Architectures. The rule from that world — never fix a running thing, replace it — now applies to application code: when rewriting is cheap, editing in place accumulates entropy.

His "deletion test" asks engineers to imagine deleting an entire implementation. What people usually mean by "we can't throw the code away," she writes, is that they don't know which behaviours are required, which failures are unacceptable, which invariants must hold, how to tell whether a new version is correct, or which bugs are deliberate fixes for forgotten edge cases. "Those are not code problems. They are evaluation problems." Code becomes precious when it is the only place knowledge lives; otherwise it acts as a cache — useful while current, disposable when stale.

Where the discipline goes

The practical consequence, she argues, is that attention should move from lines of code to other artifacts: architecture, behavioural and characterization tests, capture and replay, traffic splitting and observability. Production, in her formulation, is not what happens after development ends — production is a stage of development.

Majors is blunt that humans are the weakest link at validation — repetition and nitpicking are exactly what people are bad at — while remaining valuable for creativity and leaps of logic. Her prediction for 2026 is a return to discipline: nondeterministic systems in production will require more engineering discipline, not less. She also notes that only around 5% of engineering teams, and certainly fewer than 10%, work in the short, fast feedback loops she considers the cardinal sign of that discipline.