Agent Action Authorization Research Note

Agent Action Authorization Research Note

Meeting-derived action should move through explicit states: candidate, confirmed commitment, assigned owner, approved write, executed change, verified result, and recoverable record.

A transcript can be wrong. A speaker can brainstorm without committing. A participant can commit without authority to bind another person. A correct commitment can still belong in the wrong system.

NIST's AI RMF Measure playbook recommends auditability, histories, oversight measures, override records, and documented go or no-go decisions. NIST's 2026 agent identity and authorization concept paper frames agent identity, delegation, least privilege, auditing, and human-in-the-loop authorization as open enterprise design problems.

The smallest safe write uses a named human or service identity, minimum scope, a complete preview, idempotency or duplicate protection, a returned record identifier, and a reversal or correction path.

Human review must occur at the authority transition, not merely somewhere in the workflow.