Live-SWE-agent is a software engineering agent powered by LLMs that can autonomously diagnose, write, and fix code across real-world issues. The agent dynamically modifies and extends its own behavior during runtime rather than relying on pre-defined static heuristics. It achieves strong results on SWE-bench Verified benchmarks using models like Claude Opus 4.5 and Gemini 3 Pro, demonstrating how modern LLM-based agents can build self-improving systems that continuously evolve through live execution.