What happens when a tool makes you faster at producing work you wouldn't have wanted to produce? That's the question developer Erik Husom poses in a close look at his experience with AI agents across writing, research, and software development.
His core argument: speed is not the same as value. Completing tasks faster doesn't automatically produce better outcomes. The efficiency gains that dominate most AI coverage don't map cleanly onto meaningful progress - and Husom is specific about where that gap shows up.
Writing and Research: The Cases That Don't Hold Up
On writing, Husom found AI feedback pulled his prose toward outputs he wouldn't have chosen independently. Not wrong - just not his. For research work, AI agents produced summaries that read as complete and coherent but lacked genuine insight. They identified the center of existing thought rather than advancing beyond it.
This is an underreported failure mode. The problem with AI-assisted research isn't inaccuracy - models have gotten quite good at accurate synthesis. The problem is that original thinking requires going somewhere other than the middle, and a model trained to aggregate existing sources won't do that.
Coding Is the Exception
Software development is where Husom draws a different conclusion. He concedes that coding agents have improved substantially since fall 2025 and that avoiding them entirely would put developers at a real professional disadvantage. His own approach is disciplined: limit requests to one or two files at a time, maintain ownership of design decisions, never hand over architectural choices.
That tracks with what experienced developers broadly report. Coding agents perform best when tasks are well-defined and tightly scoped. They create problems when given too much latitude - and untangling the results costs more time than the original shortcut saved.
The Arguments Most AI Coverage Ignores
Husom raises three issues that don't get much attention in mainstream coverage: copyright exposure from training on scraped content, the energy cost of running large inference workloads at scale, and the displacement of workers whose outputs trained these systems. He's making these points not as someone who rejects AI categorically but as a practitioner who uses the tools and finds specific value in some of them.
His conclusion - slower, more deliberate development - won't change industry direction. But the underlying question is legitimate. If the primary gain from AI agents is doing existing work faster, and faster produces more output rather than better output, the actual progress is narrower than the conversation suggests.
For daily users, the coding applications hold up. But Husom is right that "faster" is doing a lot of work in agent conversations - and there's a real gap between completing tasks more quickly and producing outcomes worth completing.