Claude Code generates code — but who builds the architecture?

Claude Code is impressive. In minutes, it generates REST controllers, services, JPA entities, React components. But one question remains: who decides the structure of all this?

The problem with unframed vibe coding

Ask Claude Code to generate an inventory management application. It will create files, classes, endpoints. But:

• What naming strategy? What package organization?
• Are migrations versioned?
• Are there integration tests or just unit tests?
• Is Spring Security correctly configured?
• Does the API respect REST conventions?

Claude Code responds to what you ask. It doesn't decide the architecture for you — and if you don't specify it, it invents its own each time.

Architectural drift in client projects

In client projects, drift happens fast. After a few weeks of unframed vibe coding:

• Naming conventions differ between modules
• Tests are missing or superficial
• Migrations aren't reproducible
• No developer on the team truly understands the whole picture

This isn't an AI competence problem — it's an absence of framework.

What Katappult solves

Katappult is an AI agent supervision framework. Before Claude Code generates anything, the architecture is defined, conventions are set, technical constraints are specified.

AI agents operate within this framework — they accelerate delivery without creating technical debt.

Concrete result

• Consistent architecture throughout the project lifecycle
• Code readable and maintainable by any developer
• Tests generated according to best practices
• Versioned, reproducible Liquibase migrations
• Security wired in from the start

The right posture

Claude Code is an exceptional tool. But a tool needs a framework. The role of a senior developer in 2026 isn't to code — it's to define the framework within which AI codes on their behalf.