Vibe Architects: How Agentic AI Redefines Coding

Vibe Architects: When Agentic AI Redefines Coding

Meet the Vibe Architects

The primary goal of our study was to understand how individuals who are not professional developers use agentic AI. The study was inspired by Claude Cowork's explicit focus on non-coders. We wanted to know how well Cowork serves this population.

For the study, we aimed to recruit individuals who were not developers, did not work in our field (digital design, product, or UX), and used open agentic tools like Claude Cowork or Claude Code. We managed to find 7 non-technical individuals, but we struggled to find people outside our sector (which may indicate how rare these individuals are).

Our participants included:

• An operations specialist building automation pipelines and deploying web applications for her team
• A product designer at a financial institution creating personal software tools during her evenings and weekends
• A marketing startup founder managing a headless orchestration system with proactive suggestions and multi-agent coordination
• A product manager who put his entire team on a Claude-based "operating system" that replaced meetings, shared information, and facilitated decision-making

All of these individuals were building complex agentic systems using Claude (along with other AI tools) for a wide range of personal and professional purposes — without any technical training.

How the Vibe Architects Build

Next-generation models are extremely effective at interpreting ambiguous prompts. As a result, many participants reported that they did not engage in prompt engineering at all — they provided stream-of-consciousness instructions to their AI systems, often using dictation apps like Wispr Flow.

However, there is still a bit of "engineering" that takes place. This concerns less the choice of words in the prompt and more the selection of tools, techniques, and approaches to achieve what they want. This process often involves asking Claude himself for help.

For example, one participant reported that he always asked Claude: "What would make this world-class?" He said that Claude often responded with suggestions for improvement, and he almost always accepted those suggestions.

This was a common theme among our sample — our participants often delegated decision-making, as well as execution, to Claude. They exercised minimal control over what Claude did.

How the Vibe Architects Learn

Our participants reported that they developed their skills as vibe architects in two main ways: by spending a lot of time experimenting and by learning from a community.

Time Spent Experimenting

Each of our participants invested considerable time getting familiar with tools like Claude Cowork and Code. One participant reported spending 8 hours a day, 6 to 7 days a week in AI discussions. Another estimated spending 4 to 5 hours a day between Claude, Codex, and other tools. A third experimented with these tools on his personal computer during his evenings and weekends, as his employer restricted AI use at work.

This time investment does not stabilize either. These tools continue to evolve as AI labs release model improvements and product updates, and other vibe architects share their insights, often forcing vibe architects to engage in continuous reevaluation.

Learning from Other Vibe Architects

Across the seven participants, AI coding products consistently failed as a source of learning, regardless of the participant's technical level. This makes sense: LLMs are inherently opaque and uneven, and users do not see their limitations until they are confronted with them. Even when working with products from a single AI lab, like Anthropic's, different models can behave differently and vary in their ability to handle specific tasks.

Although Cowork was designed to help non-technical individuals get started with agentic AI, most of our participants did not really understand its relationship with Code (even those who sometimes used both).