Periodic Agent-Control Review: An Evidence Workbook

What evidence should a periodic agent-control review retain?

Retain the review scope, control version, sample-selection method, relevant action records, supporting approvals, test results, exceptions and follow-up decisions. Collect enough to reproduce the conclusion without copying unnecessary sensitive content. Distinguish a designed control, a successful test and observed operation during the review period.

Build a review packet with a defined boundary

State the workflow, period and control objective before selecting evidence. A review of approval matching should not silently become a claim about legal accuracy, data security or all agent behavior. Identify the population from which any sample was drawn.

NIST SP 800-53A provides a methodology for assessing security and privacy controls. [S1] The workbook below is a narrower editorial adaptation for an agent workflow, not a reproduction of that assessment program.

• Scope: workflow and review period.
• Control version: configuration and relevant change dates.
• Population: how many eligible events existed and how they were identified.
• Sample: selection method and excluded events.
• Evidence: references to relevant records, with access appropriate to their sensitivity.
• Findings: exceptions, consequence, owner and due date.

Synthetic ten-event workbook

Assume a fictional review of ten agent-generated reports, each requiring approval before distribution. Eight have a retrievable approval record. Of those eight checkable events, seven match the approved artifact and destination; one shows a changed recipient. The other two cannot be assessed for conformity from the retained evidence. These are invented teaching data, not measured operations.

Record missing evidence separately from a confirmed mismatch. A missing approval record does not automatically prove that no approval occurred; it means the review cannot establish it from the retained evidence.

• Approval evidence available: 8 of 10 events, or 80% of this sample.
• Artifact/destination conformity: 7 of 8 checkable events, or 87.5%; 2 of the 10 sampled events remain unassessable.
• Exception E-03: approval record unavailable; obtain or reconcile evidence.
• Exception E-07: approval record unavailable; outcome remains unresolved.
• Exception E-09: recipient differs from the approved recipient; investigate cause and affected distribution.

Do not overread the percentages

The two measures answer different questions and may overlap. They should not be added together or averaged into a compliance score. A single consequential mismatch may warrant action even if most sampled events passed.

Record how the ten events were selected. A convenience sample illustrates those events; it does not estimate the population's failure rate without a suitable sampling design. A system's own success label is evidence of its report, not independent proof of the underlying outcome.

• Evidence completeness: can the reviewer trace the required record?
• Control conformity: did the event meet the defined rule?
• Impact: what consequence followed from the exception?
• Coverage: which periods, tools or event types were not assessed?

Close a finding with a correction and a check

A useful finding names the defect, cause, corrective action, owner and verification result. If a recipient change bypassed review, a correction might bind approval to the destination and require a new decision when it changes. The follow-up test should exercise that specific failure.

NIST's control catalog distinguishes control functionality from assurance about it. [S2] A saved setting supports a configuration claim; an executed test and observed result support a different, narrower claim about behavior.

• Finding: recipient change was not compared with approval.
• Correction: add destination comparison to the gate.
• Check: changed-recipient fixture is denied; approved fixture still succeeds.
• Closure: identify the tested version and remaining limits.
• Retention: follow the applicable policy and obligations; do not invent a universal period.

Scope and limitations

Proposed operational workbook. It is not a mandated retention schedule, a formal audit opinion or a real organization's control assessment.

All counts and findings are synthetic. The workbook does not establish compliance, prescribe a retention period or provide statistical assurance about any real population.

Sources

1. NIST SP 800-53A Rev. 5: Assessing Security and Privacy Controls — Control-assessment methodology; publication page identifies later release updates. Checked 2026-11-05
2. NIST SP 800-53 Rev. 5: Security and Privacy Controls — Flexible security/privacy control catalog; source warns that framework crosswalks do not establish equivalence. Checked 2026-11-05

Explore further

Use this evidence workbook when thinking about the workflow and human-review questions associated with Agentic Attorney at https://agenticattorney.com/. It does not report an audit of that project. Explore Agentic Attorney.