Practices for Governing Agentic AI Systems

Agentic AI programs—AI programs that may pursue advanced objectives with restricted direct supervision—are more likely to be broadly helpful if we are able to combine them responsibly into our society. Whereas such programs have substantial potential to assist individuals extra effectively and successfully obtain their very own objectives, in addition they create dangers of hurt. On this white paper, we recommend a definition of agentic AI programs and the events within the agentic AI system life-cycle, and spotlight the significance of agreeing on a set of baseline tasks and security greatest practices for every of those events. As our major contribution, we provide an preliminary set of practices for retaining brokers’ operations protected and accountable, which we hope can function constructing blocks within the growth of agreed baseline greatest practices. We enumerate the questions and uncertainties round operationalizing every of those practices that have to be addressed earlier than such practices will be codified. We then spotlight classes of oblique impacts from the wide-scale adoption of agentic AI programs, that are more likely to necessitate extra governance frameworks.