The code base (and libraries and packages) serve as a knowledge base of definitions (what things are) and procedures (what things exist). The symbolic AI engine can then take information from this (essentially code snippets) and combine them in novel ways. Example: If you know what a RESTful API is (how it is defined), what MongoDB is and which data models you want to deal with, you can generate the source code for an API handling those data models with MongoDB backend from this. Other example: If you know what requests your API supports, what the semantic data types of the objects and fields are and what a UI-based application is, then you can generate a web, mobile or desktop application from this providing a wrapper around this API.
The goal here is not to replace programmers entirely, but to alleviate the pain of having to code things up manually, especially boilerplate. Because you can deal with issues at a higher level of abstraction now, this can make programming more approachable to (so far) non-programmers.