Software Self-Extension with SelfEvolve: an Agentic Architecture for Runtime Code Generation

A team of researchers including Md Asif Iqbal Fahim, Oluwadamilola Adebayo, and Alessio Ferrari has introduced SelfEvolve, a novel agentic architecture that enables software systems to autonomously extend their own capabilities during runtime. Unlike traditional self-adaptive systems that merely reconfigure existing components, SelfEvolve uses large language models (LLMs) to generate entirely new functional modules in response to user requests. This process, called 'self-extension,' allows software to evolve beyond its original design without requiring system restarts or developer intervention.

The architecture was rigorously evaluated across 11 distinct self-extension tasks, where it achieved an impressive average Pass@1 score of 92.7% (51 out of 55 tasks completed successfully). This performance represents a 61.8% improvement over the best baseline system, AutoGen, with statistical significance. The system also outperformed other developer-focused code generation frameworks like MetaGPT and AgentCoder, demonstrating that agentic orchestration specifically designed for runtime operation yields superior results compared to tools built for offline development.

This research, accepted at the 21st International Conference on Software Engineering for Adaptive and Self-Managing Systems, provides preliminary evidence for a new paradigm where systems can autonomously evolve to meet user needs. The work bridges the gap between static code generation and dynamic system adaptation, paving the way for more individualized and continuously self-improving software that can respond to unforeseen requirements in real-time.