David Heinemeier Hansson (@dhh), the creator of Ruby on Rails and co-owner and CTO of 37signals, said in an August 11th post on X that working with AI has restored a level of creative urgency he had not felt since his earliest days with Ruby. "A million ideas, all bouncing around in my head, just waiting to be prompted," Hansson wrote.
"A million ideas, all bouncing around in my head, just waiting to be prompted"
His shift matters beyond software development. It offers a preview of what is likely to happen in journalism once AI systems can reliably take an assignment, gather material, produce a draft and return it for review: The scarce work moves from typing every sentence to choosing the assignment, directing the process and deciding what is fit to publish.
That transition will not eliminate reporting, editing or accountability. Hansson's own experience suggests the opposite. As AI takes on more production work, professional judgment becomes the constraint. For developers, that means architecture, review and product selection. For journalists, it means sourcing, verification, framing and editorial responsibility.
Hansson changed his mind when AI could do assignments
Hansson's reversal unfolded as coding tools moved beyond autocomplete. In a December 2024 essay, he compared AI to a knowledgeable junior programmer whose work often contained bad assumptions, unnecessary dependencies and architectural dead ends. Code generated inside familiar Ruby and JavaScript domains, he wrote, generally required material rework.
The break came when agents gained terminal access, test execution, web searches and other tools. In a January 7th, 2026 post, Hansson said those capabilities allowed agents to contribute production-grade code under supervision. Instead of trying to finish the sentence he was typing, an agent could take a separate assignment, work through a codebase and return a result for review.
By April, Hansson described an agent-first workflow in an interview with The Pragmatic Engineer. He ran multiple models in separate terminal panes and spent much of his time reviewing their output instead of manually entering each line. He still rejected claims that agents were writing more than 90% of production code when quality and cohesion mattered.
That qualification is central to the lesson for journalism. The consequential change is not that an AI system can generate text. It is that a professional can delegate a bounded assignment, inspect the result and run more experiments than manual production allowed. Hansson's August post moves the case beyond efficiency: Lowering the cost of testing an idea changed his desire to create.
Journalism will face the same workflow shift
The comparison is not between code and copy as interchangeable products. It is between two forms of knowledge work in which practitioners have often treated manual production as inseparable from professional identity. Hansson initially saw AI assistance as an interruption to programming. He embraced it when the tools became capable enough to operate under his direction without requiring him to surrender final judgment.
Journalism is likely to follow the same path. Resistance will persist while AI produces generic prose, weak assumptions or unverifiable claims. Adoption will accelerate when agents can handle coherent parts of an assignment while leaving journalists responsible for the source relationships, verification and decisions that make the work credible.
Hansson's experience also cautions against reducing that future to a productivity statistic. His August 11th post offered no data on time saved, code shipped, projects attempted or the proportion of agent-generated work that survived review. The stronger claim is about creative capacity: Prompting lets him pursue more ideas while concentrating his attention on direction and selection.
For a journalist, the analogous gain would not be publishing unreviewed machine output. It would be testing more reporting paths, interrogating more material and developing more possible frames before deciding which story deserves publication. The bottleneck shifts away from producing a first draft and toward exercising judgment.
37signals is building around delegated work
Hansson has already translated that conviction into product and infrastructure decisions at 37signals, which makes Basecamp and HEY. In March, he said the company had revamped Basecamp's API, released a command-line interface and published instructions for agents to operate the project-management service.
An agent can search projects, summarize activity, create to-do lists, post updates, upload files and arrange schedules. Rather than insert a generic chatbot into Basecamp, 37signals is preparing the service to be controlled by agents customers choose. Hansson said the company had tested native AI features over 18 months and declined to ship them because the results were not useful enough.
That is the strategic pattern journalism should watch. AI does not need to replace the visible product to reorganize how the work behind it gets done. APIs, tools and assignment instructions can turn existing systems into infrastructure for agent-supervised workflows while humans retain control over the final output.
Hansson's conversion does not prove that every AI-assisted experiment will be good, or that supervision can remove every error. It shows how quickly skepticism can collapse once delegation becomes useful. Journalism's standards make the review burden different and the consequences of failure more serious, but they do not exempt the profession from the same economic and creative pressure. Once agents can perform bounded editorial assignments well enough to expand what a journalist can investigate and test, the shift from drafting everything to directing more of it becomes a matter of time.