AI has made it much easier to go from an idea to a working application.
You can describe an application, generate an interface, connect a database, and have something running surprisingly quickly.
But there is another problem that gets less attention:
hat happens after the application already exists?
That's where building software with AI becomes a different problem.
When you're starting from scratch, the AI has a relatively simple job. It can generate a structure, create components, set up routes, and build functionality around a new idea.
An existing product is different.
There is already a codebase.
There are existing components, conventions, APIs, authentication flows, database relationships, dependencies, and decisions that were made months or even years ago.
Adding one feature means working within all of that.
The real challenge isn't generating code
Imagine an agency has already built a SaaS application for a client.
The client now asks for:
"Add team invitations and role-based permissions."
That sounds like one feature.
In reality, the developer may need to understand:
how users are stored
how authentication works
where permissions are handled
how the frontend communicates with the backend
which components are already available
how routes are structured
how the database is organized
what existing functionality could be affected
Generating code is only one part of the job.
Understanding where that code belongs is often the harder part.
Existing code changes the AI development problem
This is why I think there is an important distinction between:
"Build me an application."
and
"Understand my application and help me develop its next feature."
The second problem requires context.
An AI system working with an existing repository needs to understand more than individual files. It needs to understand relationships between those files and the conventions used throughout the project.
*For example, before changing a dashboard component, it may need to know:
*
which component is actually responsible for the UI
which design system the application uses
where the related API lives
how state is handled
whether a similar component already exists
what dependencies the project expects
Without that context, AI-generated changes can easily become isolated implementations rather than changes that fit the application.
Feature development is an ongoing process
Real software development rarely ends after the first release.
A product might go through:
Idea → MVP → customer feedback → new feature → bug fix → refactor → integration → testing → deployment
Then the cycle starts again.
That means an AI development workflow needs to support more than initial generation.
It needs to work across the lifecycle of a product.
This is especially important for teams that are continuously developing an existing application rather than creating a new project every time.
Where AI teams become interesting
Another part of this problem is that software development is not really one task.
A feature can involve planning, implementation, frontend work, backend work, testing, verification, and iteration.
That raises an interesting question:
Could multiple AI agents work together like a development team instead of treating every prompt as an isolated coding request?
For example:
One agent could help break a requirement into features.
Another could work on implementation.
Another could review or test the changes.
The important part isn't simply having more agents.
The important part is whether those agents share enough context about the actual project to work on the same application coherently.
The next stage of AI-assisted development
I think the next stage of AI-assisted development will be less about:
"AI can write code."
We already know it can.
The more interesting question is:
"Can AI understand an existing software product well enough to help develop it continuously?"
That changes the requirements considerably.
You need project context.
You need architectural understanding.
You need awareness of existing components and conventions.
You need a way to make incremental changes.
And you need verification rather than simply accepting whatever code was generated.
That's the problem space I'm currently interested in while working on Shipd.
We're exploring how AI can be used not only to build applications from scratch, but also to help teams work on real applications, existing codebases, and ongoing feature development.
There is still a lot to figure out.
But I think that is where the conversation around AI development is becoming more interesting.
Building the first version is one challenge.
Continuing to build the tenth, twentieth, and hundredth feature is another.
Top comments (0)