Orchestrated AI Dev Team
Vibecoding tools generate new applications. This framework was built for the harder case: driving AI-assisted development inside a real, existing codebase while keeping constraints, review and auditability intact. An Architect agent reads the spec and the current code, then emits a machine-parseable contract of file-level constraints that the downstream agents must work within. Backend, Frontend, QA and Security agents execute in sequence behind a deterministic verifier that blocks edits to protected files.
The Problem
AI coding tools are optimised for greenfield generation. Pointed at a large existing codebase they edit files they should not touch, skip review, and leave no trace of what changed or why.
The Approach
Five specialised agents run in sequence: Architect, Backend, Frontend, QA and Security. Role-Based Model Assignment puts Opus on the Architect, where reasoning quality decides the whole run, and Sonnet on the executors, where cost dominates. A deterministic verifier protects package.json, .env and tsconfig.json. State persists in project-state.json and every step is git-diff verified.
Outcomes & Impact
Technologies Used
Interested in similar work?
I'd love to discuss how I can help with your project or organization.