Organizations that deploy AI agents often need to approve changes—such as adding a stricter safety gate or swapping in a new model—before those changes go live. A new paper on arXiv proposes a framework called agent-risk for assessing the risk of such changes, starting from the logs the agent has already produced.

The approach extends classical risk analysis to handle two complications common in real deployments: records that are only partial, and changes that affect only part of a running agent. Rather than assuming full observability or a clean before-and-after comparison, the framework works with the incomplete information available.

The abstract is brief and does not include experimental results or implementation details. It frames the problem as an approval process and positions the contribution as an adaptation of established risk-analysis methods to the agent setting.