Engineers Must Still Write Code, Not Just Build Agents
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- The author reframes the engineer's job as building the "software factory" — prompts (skills, AGENTS.md, knowledge bases) and reactive guards (tests, linting, type systems, evals) — rather than shipping software that agents generate from English instructions.
- Writing code remains useful, the author argues, because English is "an under-specified language" for expressing computation, and thinking directly in the execution environment produces better architectural judgment than proxying through prompts.
- Hands-on coding protects against "slop": when engineers only read and approve AI-generated changes, the author says they lose ownership, micro-adjustments become harder, and accumulated fragility hurts both reviewers and downstream agent runs.
- Agents amplify bad human decisions because they bias toward minimal, safe changes rather than cleanup; the author cites a case where one pre-caffeinated comment about using browser local storage led to wrapping and indirection that "probably tripled the LoC."
- The essay characterizes coding agents as "freshly onboarded interns," not deterministic compilers — they read partial, possibly slopified code and act on imprecise change descriptions, so humans cannot surrender thinking and taste to them.
- Cleanup through code, not prompts, paid off: the author says joyfully deleting code and exploring it directly led to a better architecture than trying to direct the same changes through English prompts alone.
- Agents are unlikely to follow the boy scout rule, the author argues, because they optimize for the current change being safe; this makes human-driven architectural cleanup a necessary complement rather than a redundant luxury.
Why it matters: For engineering teams racing to delegate work to AI agents, the essay argues the fastest path is hybrid: humans must still write and edit code to catch fragility, enforce consistent architectural principles, and prevent the kind of defensive wrapping that "freshly onboarded interns" naturally produce around human mistakes.
