Agentic artificial intelligence for interactive and incremental software development

One or more instances of non-transitory computer-readable media collectively having contents configured to cause a computing system to perform a method comprising:

• receiving, from one or more users, a natural language description of at least one of task, purpose, or function of a target software;
• obtaining initial code for the target software;
• outputting, using one or more trained machine learning models, cumulative modifications to the initial code that correspond to a plurality of intermediate states of development toward final code for the target software, wherein for each intermediate state of at least a subset of the plurality of intermediate states:
• obtaining, from the one or more users, interactions with a cumulative modification corresponding to the intermediate state;
• and
• generating subsequent cumulative modifications based, at least in part, on a revised intermediate state in accordance with the interactions;
• outputting the final code for the target software;
• and
• updating the one or more trained machine learning model based, at least in part, on the final code and the interactions obtained with the cumulative modifications.

A system comprising:

• a computing device configured to:
• receive, from one or more users, a natural language description of at least one of task, purpose, or function of a target software;
• obtain initial development files for the target software;
• output, using one or more trained machine learning models, cumulative modifications to the initial development files that correspond to a plurality of intermediate states of development toward final development files for the target software, wherein for each intermediate state of at least a subset of the plurality of intermediate states:
• obtain, from the one or more users, interactions with a cumulative modification corresponding to the intermediate state;
• and
• generate subsequent cumulative modifications based, at least in part, on a revised intermediate state in accordance with the interactions;
• output the final development files for the target software;
• and
• update the one or more trained machine learning model based, at least in part, on the final development files and the interactions obtained with the cumulative modifications.