Most tech coverage treats artificial intelligence as a marketing buzzword or a feature to be bolted on; this piece argues that a mature, five-year-old startup has fundamentally rewritten its operating system in a matter of weeks. Gergely Orosz provides rare, granular access to how a non-AI-native company with 50,000 paying customers pivoted from skepticism to total dependency, revealing that the real revolution isn't in the models themselves, but in the interface that makes them usable for non-engineers.
The End of the "Copilot" Experiment
Orosz frames the narrative not as a sudden epiphany, but as a three-year journey of disciplined failure. He details how Craft, a document editor known for its sleek design, spent years resisting the industry pressure to ship gimmicky AI features. "We knew we had one shot at showing users that AI is useful for them," Balint Orosz, the founder, is quoted as saying. "And until we had a 'wow' moment ourselves, we would not go all-in on AI. We didn't want Craft to have the same negative associations with AI as Copilot has." This restraint is notable; while competitors rushed to integrate chatbots that often produced hallucinations or shallow summaries, Craft's leadership waited for the technology to mature enough to solve actual problems rather than just creating noise.
The turning point arrived not with a new algorithm, but with a shift in capability. Orosz describes how the founder attempted to build a complex shape-recognition feature, a task estimated to take weeks. Using a reasoning model, he completed it in a single day. "It was the first time Balint had to admit they couldn't have shipped a feature without AI," Orosz writes. This admission is critical because it shifts the conversation from "AI as a helper" to "AI as a prerequisite for execution." Critics might argue that this anecdote is cherry-picked and doesn't represent the reality for most engineering teams, yet the subsequent organizational shift suggests this was a genuine inflection point for the company.
We knew we had one shot at showing users that AI is useful for them. And until we had a 'wow' moment ourselves, we would not go all-in on AI.
From Terminal to "Visual Claude Code"
The most distinctive contribution of this piece is the description of how the company moved beyond the command-line interface. Even after building a powerful terminal-based agent, the team realized that non-technical staff—customer support, marketing, and HR—were alienated by the text-only environment. Orosz notes that the terminal felt "locked away" to users who weren't developers. In a bold move, the founder spent his Christmas break building a graphical user interface on top of the agent framework, a task he completed in two weeks using the very AI tools he was trying to promote.
The result was "Craft Agents," a tool that transformed the agent experience into something resembling an email client or a Slack workspace. Orosz highlights the surprising outcome: "Non-technical users of Craft were struggling with a terminal experience, so Balint imagined what Claude Code with a UI and strong opinions would look like." This design-first approach allowed customer support to build custom workflows for bug triaging and education verification without writing a single line of code. The article suggests that the barrier to AI adoption isn't the intelligence of the model, but the usability of the interface. If the tool requires a terminal, it remains a developer toy; if it feels like a familiar app, it becomes a company-wide utility.
The New Software Development Lifecycle
Perhaps the most unsettling implication for the industry is the shift in how software is actually built. Orosz describes a culture where the traditional safety net of code reviews and pull requests is being abandoned in favor of speed and iteration. "Difficult migrations take a week instead of months," he reports, noting that some developers are struggling with the pace of change, while others have quit. The new workflow involves "weaving in" ideas from AI agents rather than strictly reviewing their code, a fundamental break from the collaborative engineering practices that have dominated for decades.
This rapid iteration brings significant risks. Orosz admits that "some devs struggle with fast change, and even quit," hinting at a potential talent churn as the industry adjusts to a model where human oversight is reduced. The article posits that "enterprise-only capabilities could become standard consumer-grade offerings," which could destabilize the market for specialized B2B software. While this promises incredible efficiency, it also raises questions about quality control and the long-term maintainability of codebases built at this velocity. The author suggests that the "death of pull requests for open source" and the rise of "remixing" are inevitable trends, but the human cost of this transition remains a point of tension.
Bottom Line
Orosz's analysis is compelling because it moves beyond the hype cycle to document a concrete, operational transformation in a real company. The strongest part of the argument is the demonstration that the interface, not just the intelligence, is the bottleneck for AI adoption. However, the piece glosses over the long-term risks of abandoning code reviews and the potential for a fragmented workforce where only those who can adapt to AI-driven speeds survive. The industry should watch closely to see if this "visual agent" model scales beyond a single startup or if the lack of traditional engineering rigor leads to systemic failures down the line.