Seen this pattern repeatedly building a shell plugin with Claude. It defaults to writing everything from first principles rather than reaching for existing tools. 200 lines of custom YAML parsing when a one-liner would do. Adding "always check if a library or existing tool solves this before writing custom code" to CLAUDE.md cut this down significantly.
The strongest signal I have seen for whether AI actually reduces maintenance cost is whether the developer treats AI output as a first draft or a final artifact.
When I use AI tools on existing codebases - understanding unfamiliar modules, generating targeted refactors, writing migration scripts - the maintenance burden genuinely drops. The AI is working on code I already understand architecturally, so I can evaluate its output quickly.
The problem shows up when AI generates greenfield code that nobody deeply understands. That code still has to be maintained by humans who did not write it AND did not design it. At least with code another human wrote, you can reason about their intent from naming, structure, and commit history. AI-generated code often lacks that legibility because the "author" had no persistent intent across files.
The article is right that we need to measure maintenance cost, not just velocity. In practice that means tracking time-to-understand and change-failure-rate on AI-assisted code vs. human-written code over months, not days.
One thing I'd add from building a shell plugin that routes natural language to agents: the detection heuristic matters way more than flag conventions.
We spent a lot of time on when to run something as a shell command vs send it to an LLM. The hard lesson: false positives are much worse than false negatives. "git push --force" accidentally going to an LLM instead of executing is the kind of thing that kills user trust instantly. Our heuristic ended up very conservative.
The bigger surprise was the real-time visual indicator. We added a small color signal showing "this goes to shell" vs "this goes to agent" as you type, and it changed how people wrote more than anything else. Before it, people hedged natural language queries with shell-like syntax just in case. After it, they wrote normally.
On the isatty point — right for automation. But there's a third mode worth thinking about: "orchestrated interactive," where a human is watching the agent use your CLI and needs to step in. Pure non-interactive breaks that entirely.