Developers are attached to tools because tools encode trust

There's a reason developers get weirdly emotional about their editors. It's not nostalgia, it's trust built through repetition. Spend a decade in Vim and your fingers know what to do before your brain finishes the thought. That's not stubbornness, that's what The Pragmatic Programmer called a "sharp tool," one that becomes an extension of your hand. Swap it out and you're not just learning new keystrokes, you're relearning how you think about writing code.

Agentic coding tools blow that model up. Instead of precise, predictable inputs, you're typing ambiguous English at a system that outputs entire applications in seconds. Bjarne Stroustrup put it bluntly: English is a lousy language for anything that needs to be unambiguous. And the data backs up the unease — TLDR's own Developer Survey found AI usage climbing from 76% to 84%, while trust in its output dropped from 40% to 29% over the same stretch. People are using it more and believing it less, which is a strange place for any tool to sit.

The deeper problem isn't the AI itself, it's that the tools around it were built for a slower, more deterministic process. Linters, CI/CD pipelines, story-point estimates — none of that machinery assumed a tool could rewrite hundreds of lines in a blink. Code review, once a minor checkpoint, is now the bottleneck, because reviewing a 300-line AI diff takes real cognitive effort even when the code technically works. Running that code isn't free either — compute, dependencies, and the ever-present cost of failure don't disappear just because generation got cheap.

What's emerging as the fix isn't better AI, it's tighter human accountability wrapped around it. Charity Majors' line about "human-in-the-loop" being a pity invite says it well: whoever pushes the commit owns it, agent or no agent. Teams are experimenting with shared prompting sessions, PRs that include the AI conversation transcript, and painstakingly explicit spec files, because as Scott Hanselman noted, anything left unspecified in a prompt just gets left to chance. There's also a quieter discipline emerging — reusing components instead of regenerating them, and simply not using AI where a boring deterministic script has worked fine for six years, per Anil Dash's blunt assessment.

None of this fixes the trust gap by itself. Trust was never really about the tool — it was about predictability, about knowing Monday's experience would match Tuesday's. AI breaks that assumption constantly. But process discipline, explicit specs, and clear ownership can rebuild some of what got lost, even if the tools underneath keep shifting shape.