Self-evolving agents are systems that modify their own harness—the code or scaffolding that defines how they operate—and then need a way to judge whether each rewrite is an improvement. The authors of a new arXiv paper argue that the signal for this decision should be the check a rewrite has to pass, and they examine test suites as that signal.

Verifying a rewrite by running a full agent rollout is expensive and subjective, the paper notes. Executable test suites, by contrast, can serve as a compact and objective improvement signal. The title's "small but mighty" framing suggests that even modest test suites may be enough to guide self-evolution without the cost of extensive rollouts.

Because the source is a single arXiv abstract, this digest does not compare findings across studies. The full paper would be needed to see experimental evidence for how well test suites perform against rollout-based verification.