DelegateOS: Cryptographic delegation tokens for multi-agent sys(DeepMind paper) (github.com)

Google DeepMind has introduced DelegateOS, a middleware designed to address the critical gaps in delegation and accountability within multi-agent systems. This innovative framework provides cryptographic delegation tokens (DCTs) that allow agents to delegate tasks while maintaining strict controls on capabilities and authority. Features such as monotonic attenuation ensure that when an agent delegates to another, the subsequent agent cannot expand its scope of authority beyond what is permitted. Additionally, DelegateOS includes contract-based task verification, allowing clear expectations for task completion, and attestation chains that create an auditable record of all actions taken by agents in a delegation hierarchy. The significance of DelegateOS for the AI/ML community lies in its ability to enforce trust and accountability in multi-agent environments, which have become increasingly prevalent in complex AI applications. Unlike existing frameworks that assume all agents are trusted and co-located, DelegateOS facilitates highly granular permissions without the need for inherent trust, making it particularly useful for applications such as automated code reviews or collaborative research tasks. This framework, which emphasizes cryptographic certainty in task delegation and completion, presents a solution for developers seeking to enhance security and operational integrity in their AI systems, ultimately paving the way for more robust and trustworthy agent interactions.