As AI agents move from generating answers to operating tools, a familiar login model begins to fail. A shared service account can reveal that something happened, but not which agent acted, which principal authorized it, or whether the action stayed inside its mandate.
The emerging answer is not to pretend an agent is a person. It is to give the agent a distinguishable machine identity while keeping its authority traceable to a human or organizational principal.
Identity answers “who acted?”—not “who allowed it?”
NIST’s 2026 work on software-agent identity and authorization frames agents as actors whose permissions derive from a person or organization. NIST has also highlighted the accountability gaps created when agents reuse human credentials. That distinction matters: authentication establishes the actor; delegation establishes the actor’s right to perform a particular task.
MoonVault view: an agent should possess an identity of its own, but never an unexplained authority of its own.
A complete delegation should carry context
A useful authority record should name the principal, the agent, the permitted purpose, the resources in scope, the allowed actions, the duration, the approval threshold, and the conditions for revocation. Without that context, “authorized” is too broad to be meaningful.
This is especially important across organizational and vendor boundaries. An agent may call another agent, which may invoke a third-party service. Accountability weakens each time identity and authority are collapsed into the same bearer token.
What trustworthy agent infrastructure should make possible
- Give every agent instance a distinguishable, verifiable identity.
- Bind authority to a named human or organizational principal.
- Use least privilege, short duration, explicit purpose, and task-specific credentials.
- Preserve the delegation chain across agent-to-agent and agent-to-service calls.
- Support immediate revocation, human escalation, and post-action review.
The objective is not frictionless autonomy at any cost. It is useful autonomy whose source, scope, and consequences remain understandable.
Primary sources
- NIST AI Agent Standards Initiative
- NIST: Identity and Authorization for Software Agents concept paper
- NIST NCCoE: Software and AI Agent Identity and Authorization—summary of comments