Agent platforms are starting to automate their own improvement: a deployed agent's execution traces are read, a change is derived, an automated verifier approves it, and the new version ships with no human author. That loop is the subject of a new arXiv paper, which asks how much of it can safely run unattended.
The paper's answer is framed around two preconditions: recoverability and verifier strength. Recoverability likely concerns the ability to roll back or undo an agent's changes, while verifier strength concerns how reliable the automated approval process is. Both are presented as measurable properties, not just abstract ideals.
The authors stop short of declaring the loop fully safe. Instead, they position these preconditions as necessary checks before unattended self-improvement is allowed. The abstract is brief, so details on how to measure either precondition are not included, but the framing gives platform builders a concrete starting point for risk assessment.