Vibe coding describes a way of working in which people describe software in natural language and an AI system suggests or changes code. Andrej Karpathy used the term in a February 2025 post. In a company, the approach suits experiments and bounded prototypes; anything meant to run in production should be treated like any other code.
What AI-assisted programming can do
Depending on the product, an assistant can explain, complete or edit code. GitHub Copilot offers different features for this, including a cloud agent that can work on a task in its own environment and propose changes as a pull request. Replit Agent can create an app project from a description and connect it to the platform's services. Scope and results depend on the task, configuration and available resources. Generated code can contain errors, security vulnerabilities or wrong assumptions.
A sensible place to start
A bounded prototype is a good start when you want to test an idea with users and do not need production or confidential data. Define in advance what the prototype should show, which data may be used and who reviews the results. Compare the effort with a suitable baseline from your team.
Before production use
Treat the output like any other code: run tests, review dependencies and permissions, keep secrets out of the source code and have changes reviewed by an accountable person. For production systems, architecture, operations, data protection, monitoring and recovery belong in the plan. An AI agent does not take on that responsibility.
Also check the terms of the service you use: which content is processed, which settings apply to training and retention, and which external integrations receive inputs? The answers differ by provider, product and plan.
Further reading: GitHub Copilot and its features, GitHub Copilot cloud agent, Replit Agent.

