Towards a Hybrid Code Generation Paradigm for Reliable Agent-Driven Development
14 January 2026, by Omkar Kondhalkar

Photo: base.camp
An approach to utilize documents, large language models, agents and classical code generation tools. Modern software systems demand both rapid development and strong correctness guarantees—two goals that rarely align. Addressing this tension, the work explores a hybrid code generation paradigm that combines Model-Driven Engineering (MDE) with Large Language Model (LLM)-based agents. UML Statecharts are used to formally model system lifecycles and generate verifiable finite state machines, ensuring structural correctness. LLM-powered agents are then constrained to implement complex, documentation-specific action logic, such as interactions with external APIs. The approach is demonstrated through the generation of an operator for managing resource definitions. By separating architectural rigor from contextual implementation, the study highlights how hybrid methods can balance agility and reliability in agent-driven software development.

