AI is hollowing out software engineering's middle class
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- AI coding agents have removed the natural friction that once forced teams to discuss architecture before building, with the source illustrating the shift through a scenario in which a single PR ballooned to +24,506/-3,938 lines between Friday and Monday.
- Claude is increasingly being leaned on to make design decisions engineers cannot themselves explain, with the source describing a teammate sending a long Claude conversation as justification for an architectural choice.
- Companies paying six-figure salaries in London and San Francisco are doing so for engineering judgment, not raw code production, the source argues, since implementation itself is now cheap and the LLM of the moment can already turn a spec into working code.
- Bad engineering decisions now compound faster than teams can review or unwind them, the source notes, citing the difficulty of reversing even simple AI-introduced database denormalizations while another 20,000-line PR arrives the same afternoon.
- The engineering labor market is splitting, per the source, into a smaller pool of high-judgment engineers whose value rises and a larger pool whose work is no longer scarce enough to justify premium pay, with the author extending the same prediction to most other knowledge work.
- Tools named in the scenario — including Fable (for running agents) and an "ultracode" mode — were unable to resolve a recurring bug in the team's feature, reinforcing the source's argument that AI cannot substitute for human understanding of the system.
Why it matters: For engineering managers, the piece reframes hiring as a judgment bet rather than an output bet: one mid-level engineer who can no longer justify their decisions can now generate more unreviewable code in a day than a team can untangle in a month, while top-of-market architects become more leveraged. The author extends the same logic to most knowledge work, arguing that AI compresses the time between a bad call and an unrecoverable mess.
Ask SkimNews
