Feature Execution Graphs: A Human-AI Co-Programming Paradigm for Graph-Driven LLM Code Synthesis

Abstract

Recent advances in large language models (LLMs) have enabled transformative approaches in software development, positioning Artificial Intelligence (AI) not just as an assistant but as an integral programming layer. We introduce a novel, five-tiered framework for LLM-driven code synthesis. At its base is a minimal Task Execution Meta-Language (TEML) defining atomic tasks with typed parameters, return schemas, synchronous/asynchronous and fork/join control, plus hooks for logging, security and data management. Layered atop TEML, a Domain Task Specification Language (DTSL) instantiates these primitives into semantically rich, field-specific operations and enforces valid invocation patterns. The centrepiece is the Feature Execution Graph (FXG), a directed, attributed graph whose nodes and edges encode configured tasks and their calls. A Generation Engine traverses the FXG, issues context-aware prompts to an LLM to synthesise code for each task, and packages the results either as local functions or as containerised services. Finally, an Orchestration Engine executes the synthesised pipeline by invoking tasks locally or orchestrating services in environments such as Kubernetes. Evaluated on two representative workflows, a six-node data-science pipeline and a twelve-node EEG signal-analysis pipeline, our FXG-driven approach cut manual development time by about 40%, produced code that passed unit tests on the first attempt in 90% of local runs (85% when containerised), and preserved baseline predictive accuracy while trimming up to 25% of boilerplate.