ASI-LIB-017 technical agent benchmark

Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly?

Chunqiu Steven Xia, Zhe Wang, Yan Yang, Yuxiang Wei, Lingming Zhang

Live-SWE-agent benchmark comparison charts
Figure via ar5iv rendering of arXiv:2511.13646

Live-SWE-agent is notable because it moves self-improvement into runtime. Rather than performing an offline search for a better scaffold, the agent modifies its own scaffold while solving software tasks.

Why it matters

Software engineering is the natural proving ground for recursive improvement: the agent’s substrate is code, the task output is code, and feedback can be partially automated through builds, tests, static checks, and benchmark evaluation.

ASI relevance

If live self-evolution becomes reliable, agents can adapt their tools and control flow to the problem distribution they are currently facing. That is a more fluid route to capability gains than fixed hand-authored scaffolds.