Hello everyone,
I recently published an Internet-Draft proposing a control-plane protocol for AI agents, called Agent Interaction & Delegation Protocol (AIDP):
https://datatracker.ietf.org/doc/draft-vandoulas-aidp/
The motivation comes from practical issues observed in agent systems today: implicit authorization, unsafe delegation, lack of auditability, and limited cross-domain interoperability.
AIDP proposes a standardized layer for
-
explicit intent expression,
-
capability-based authority and delegation,
-
policy-enforced execution boundaries,
-
cryptographically bound observations,
-
replay and revocation semantics.
The goal is not to introduce a new agent framework, but to explore whether agent ecosystems would benefit from a shared control-plane abstraction, similar to how other distributed systems separate control and data planes.
I’m sharing this here for technical feedback, critique, and discussion—particularly from people building or maintaining agent runtimes, orchestration frameworks, or multi-agent systems.
Best regards,
Ioannis Vandoulas