AI Generated App Architecture Research Note

AI Generated App Architecture Research Note

An AI builder must either choose an architecture or leave important decisions unresolved. The generated interface can hide both cases.

The architecture record should identify the system boundary, trust zones, data model, tenancy model, identity provider, authentication flow, authorization policy, state transitions, file storage, background jobs, external integrations, secrets, deployment units, environments, observability, backup, recovery, and exit path. It should also record which decision belongs to the builder, which belongs to a managed service, and which remains under the application owner's control.

The most expensive dependencies tend to accumulate around data and identity. A field added to a form can flow into validation, persistence, reports, exports, permissions, retention, and integrations. A tenant model can shape almost every query. An identity choice can affect account ownership, support, auditability, and migration. These are not automatically irreversible, but the cost rises as data volume, integrations, users, and operational history grow.

Base44's backend documentation currently describes a managed NoSQL, MongoDB-compatible backend with schema definitions and row-level and field-level controls. Lovable's deployment guidance describes managed, hybrid, and self-managed routes. Replit's authentication documentation distinguishes Replit Auth from a dedicated Clerk tenant and notes that the choices fit different product contexts. These differences illustrate why "full stack" does not identify one common architecture.

Portability must be reviewed at each boundary. A Git repository can preserve application code while the working system still depends on hosted identity, databases, storage, queues, deployment settings, or proprietary editor behavior. A clean-room restore is stronger evidence than a code-export screenshot.

Dalton's E098 observation that architecture is a one-door decision is directionally useful but should be qualified. Interface styling and many components remain easy to change. Data shape, authorization, tenant boundaries, and integration contracts are closer to one-way doors because later work compounds around them.

The minimum useful artifact is a one-page system context and decision record. It should show components, data ownership, trust boundaries, irreversible assumptions, alternatives considered, and the test that would prove an exit remains possible.