Sociotechnical systems in software and AI: 'Vibe coding' and the rise of the non-programmer programmer

Vibe coding is changing who builds software and how teams organise. This post looks at the sociotechnical implications, practical workflows and safeguards for teams adopting AI-driven coding—so non-programmers and developers alike can ship reliable systems.

Introduction

We are living through a subtle but meaningful shift in software practice. A cluster of tools and habits now described as "vibe coding" lets people describe desired behaviour in plain language and have an AI produce working code. Coined and brought to attention by Andrej Karpathy in early 2025, the term captures a change from tapping keys to shaping intent. That change is not purely technical: it rearranges the social, organisational and governance aspects of software work. This post walks through what vibe coding is, why it matters to sociotechnical systems, practical ways to adopt it, and the risks teams should manage.

What is vibe coding?

At its core, vibe coding is using AI development tools to generate code from natural language prompts rather than writing every line by hand. Instead of focusing on syntax and boilerplate, practitioners express the "what"—desired behaviour, data shapes, user journeys—and an AI model scaffolds functions, APIs or UI components. Several industry pieces and early research describe it as an AI‑native approach that can accelerate prototyping and broaden who can build software.

Why this matters from a sociotechnical perspective

Thinking sociotechnically means paying attention to both people and tools and how they co‑evolve. Vibe coding rearranges three familiar axes:

Roles and skills: The boundary between designer, product manager and developer blurs when prompts become a primary artefact. People who previously couldn’t code can now produce runnable prototypes, creating a new class of "non‑programmer programmer".
Workflows: The artefact that teams iterate on shifts from code files to prompts, templates and AI sessions. That changes reviews, testing practices and knowledge sharing.
Governance: Acceleration and low friction amplify the need for provenance, security checks and reproducibility. Early literature and commentary highlight both the promise and the pitfalls of this paradigm.

What the evidence says (short)

Industry writing and early academic work describe vibe coding as an emerging pattern that lowers technical barriers and speeds prototyping. Practitioners note it is particularly strong for getting concepts out of heads and onto screens quickly. At the same time, researchers urge attention to reproducibility, alignment and security when relying on generated code. The consensus: it can democratise creation but requires explicit sociotechnical practices to be safe and sustainable.

Practical guidance for teams adopting vibe coding

Below are pragmatic steps that blend social practices with technical guardrails. Treat them as a checklist you can adapt rather than a fixed recipe.

Define clear intent and ownership: Make prompts first‑class artefacts. Record who wrote or approved a prompt and what outcome it is intended to produce. That avoids the "mystery AI magic" problem when code behaves unexpectedly.
Start with scaffolding, not mission‑critical systems: Use vibe coding for prototypes, internal tools and feature flags. This aligns with observed use cases where quick proof‑of‑concepts are the biggest win.
Version control prompts and generated code together: Commit both the prompt and the AI output into the same workflow so you can trace how a change originated and reproduce it if needed.
Automate verification: Require unit tests, static analysis and dependency scans on generated artefacts. Treat AI output as you would an external library—test it, lint it and measure its behaviour.
Human‑in‑the‑loop review: Assign a code custodian—someone with engineering authority—to review and refine AI outputs before merging. Vibe coding shifts responsibility, but it doesn’t remove it.
Build prompt templates and prompt libraries: Capture reliable prompts for common patterns (CRUD, auth, data validation). Templates reduce variance and make outcomes more predictable.
Limit blast radius with feature flags and canary releases: Deploy generated features behind flags and monitor before full rollout. This is social engineering as much as technical—product owners must be comfortable rolling things back quickly.
Document AI provenance and licences: Keep metadata about model versions, tool providers and any third‑party code snippets the model might introduce. That helps with license compliance and future audits.
Invest in education: Train non‑technical creators in prompt craft, expectations and limitations. Simultaneously upskill engineers to curate and harden AI‑generated code.

When vibe coding is a good fit

Think of vibe coding as high‑leverage for ideas that need speed over absolute polish. It shines for:

Rapid prototyping and discovery, where getting a working demo changes the conversation.
Internal tooling and automations that solve immediate productivity pain points.
Design‑to‑code workflows where designers iterate on UI behaviour faster than handoff cycles allow.

Key risks and how to govern them

Early analyses stress several hazards that teams must treat explicitly:

Reproducibility: AI outputs can vary over time or across model versions. Version prompts and model metadata so you can reproduce a build.
Quality and security: Generated code may introduce subtle bugs or vulnerabilities. Use static analysis, dependency scanning and threat modelling as standard steps.
Alignment and correctness: An AI might implement something that "feels" right but fails on edge cases. Rely on tests that encode expected behaviours.
Governance and licence compliance: Be clear about third‑party code provenance. Keep legal and security teams in the loop for production use.

A simple sociotechnical workflow

Here is a compact, repeatable flow that teams can adopt right away:

Idea owner writes a short prompt that captures the desired behaviour and acceptance criteria.
Prompt is paired with a template and submitted to an AI tool to generate scaffolded code.
Engineer reviews output, writes or adjusts tests, and runs static/security scans.
Changes are merged behind a feature flag; monitoring and error budgets are defined before rollout.
Product owner validates behaviour in staging; if green, feature flag is gradually rolled out to users.

Wrapping up (with a light note)

Vibe coding is less about throwing away compilers and more about reorganising how people collaborate with software. It lowers friction for experimentation and invites new contributors into the building process. But with that invitation comes responsibility: reproducibility, security and clear ownership must travel with the speed. In short, vibe coding hands you a faster paintbrush—don’t forget to also bring a canvas, a test‑plan and an eye for detail.