The Agent Access Model

This post introduces the Agent Access Model (AAM) as an evolution of Zero Trust principles for managing software principals (agents). AAM proposes a model where agents receive short-lived, task-scoped, sender-constrained credentials. Authorization decisions are made per action against the task's accumulated state and policy, rather than trusting the agent's identity or the network. Key components include an Agent Identity Broker for issuing credentials, a Task-Scoped Access Engine for making authorization decisions based on task templates and capability ceilings, and a Mediation Layer (harness and network) for enforcing policy at the point of action. The model emphasizes making agent capabilities smaller and enforcing least privilege in real-time.