Replaying Self-Evolution's Tape: Understanding Self-Evolving Agents Requires Experimental Evolution

Inspired by experimental evolution in evolutionary biology, this position paper argues that understanding self-evolving agents requires similarly studying the evolutionary histories of agents experimentally. Current evaluation practices mainly inherit the experimental logic of optimization problems and still treat self-evolving agents as optimization systems oriented toward a fixed goal. However, history in self-evolving agents cannot simply be interpreted as an intermediate route to the final outcome. Rather, it becomes part of the system itself by retaining and reusing past history for subsequent actions. We first diagnose three forms of historical compression in current practices. We then propose an experimental evolution perspective by distinguishing three explanatory sources of evolutionary histories—adaptation, chance, and history—and further outline an operational workflow to make evolutionary histories empirically tractable. Finally, we call for a research agenda centered on this historical perspective for agent evaluation and understanding.