A new library OS substrate called Agent libOS has been proposed for large language model (LLM) agents. It separates three planes: operation admission, information-flow admission, and durable causal evidence. This separation aims to prevent authority-escalation and data-exfiltration risks by controlling what actions the agent can take based on its capabilities and permissions. The implementation provides features such as persistent processes, object memory, skills, and typed providers. According to the authors' evaluation, Agent libOS is effective in preventing safety violations and improving utility. --- Why it matters: This matters to researchers working with LLM agents because it addresses a critical issue of controlling their behavior and preventing potential security risks. The proposed solution can help ensure that these powerful models are used responsibly and safely. Source: https://arxiv.org/abs/2606.03895