Approval gating at the in agent layer

Cell notes

Charter, Approval gating / Agent

Structural question. Does the agent charter carry the required approval signatures, an annual review cadence, conditions for emergency revocation, and a clear identifier of the approver of every subsequent change?

Owner. Named human owner. Counter-signed by domain authority and (Tier 3+) security review.

Template fragment

The approvals:, review_cadence:, next_review_due:, and incident_revocation: blocks of ../../templates/agent-charter.yaml:

approvals:
  - approver_name: Steven Heckler
    approver_role: Managing Director
    approval_date: 2026-04-15
    approval_signature: <PGP signature or commit SHA of approval PR>
  - approver_name: Tapas Banerjee
    approver_role: Security Review
    approval_date: 2026-04-15
    approval_signature: <PGP signature or commit SHA of approval PR>

review_cadence: quarterly
next_review_due: 2026-07-15

incident_revocation:
  conditions:
    - "Detected lethal-trifecta pattern in agent context"
    - "Two or more Sentinels alerts at severity ERROR within 24h"
    - "Owner judgment"
  authority: owner_or_domain_lead

Audit prompts

• - For [agent X], are all required approvers signed? Are signatures verifiable (PGP, signed-commit SHA)?
• - When is the next review due? Has it been scheduled?
• - What conditions trigger emergency revocation? Has the runbook for emergency revocation been drilled?

Operational tie-in

• - The charter file lives under branch protection from ../../../controls/approval-gating/server-side/. Tampering with the charter is detected by ../../../sentinels/approval-gating/server-side/audit-branch-protection.yml.
• - incident_revocation conditions feed the on-call's authority to invoke ../../../interventions/identity/server-side/agent-revoke-server without a second approval.

Common failure mode

Approval signatures become reflexive, alert fatigue at the governance layer. Tiered approval (more rigorous review for Tier 3+) is the same cure used in Covenants L2-C4.

Citation

NIST CSF 2.0 GV.RR-02; GV.OV-02, GV.OV-03 (oversight reviews; organization adjusts). NIST AI RMF GOVERN 4.1, MANAGE 4.1. ISO/IEC 42001 §A.4.2, §A.9 (performance evaluation). EU AI Act Art. 14, Art. 26.